<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Dazbo (Darren Lester) on Medium]]></title>
        <description><![CDATA[Stories by Dazbo (Darren Lester) on Medium]]></description>
        <link>https://medium.com/@dazbo?source=rss-58a0bd8ba726------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*jFg7mJ4-yYxt5agy4MwpyQ.jpeg</url>
            <title>Stories by Dazbo (Darren Lester) on Medium</title>
            <link>https://medium.com/@dazbo?source=rss-58a0bd8ba726------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 03:57:58 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@dazbo/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com" rel="hub"/>
        <item>
            <title><![CDATA[How I Made a Voice Clone in Minutes, Using Google Antigravity, Gemini 3.8 Flash TTS and My Harness]]></title>
            <link>https://medium.com/google-cloud/how-i-made-a-voice-clone-in-minutes-using-google-antigravity-gemini-3-8-flash-tts-and-my-harness-7ffa3e426f0f?source=rss-58a0bd8ba726------2</link>
            <guid isPermaLink="false">https://medium.com/p/7ffa3e426f0f</guid>
            <category><![CDATA[google-antigravity]]></category>
            <category><![CDATA[voice-cloning]]></category>
            <category><![CDATA[tts]]></category>
            <category><![CDATA[gemini]]></category>
            <category><![CDATA[text-to-speech]]></category>
            <dc:creator><![CDATA[Dazbo (Darren Lester)]]></dc:creator>
            <pubDate>Mon, 09 Nov 2026 06:42:01 GMT</pubDate>
            <atom:updated>2026-11-09T16:12:47.401Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Leveraging my Harness: Spec-Driven Development, TDD, and Self-Correcting Loops in Action</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Qi6guiTuS0aeyoRH.png" /></figure><p>Howdy friends! Have you ever wanted to clone your own voice so you can generate expressive narration for presentations and demos, create audio versions of your blogs, or even co-host multi-speaker AI podcasts? Or maybe you just fancy having deep-fake phone calls with family members!</p><p>I certainly have! Honestly, I don’t love recording voiceovers — mostly because fluffing a line and having to go back and correct your mistakes is a pain.</p><p>In this post, I’ll show you how I used <strong>Google Antigravity</strong> and my reusable agent harness to build <a href="https://github.com/derailed-dash/voice-replication">voice-replication</a> — an open-source Python application that clones your voice and generates expressive single- and multi-speaker speech with <strong>Gemini 3.8 Flash TTS</strong>. And if you want to clone your own voice (or script multi-speaker dialogues) without writing any code yourself, you can simply clone the repo, record two short WAV clips, and run the steps we&#39;ll walk through below!</p><p>Google first introduced custom voices and voice cloning in its “Chirp 3” Text-to-Speech (TTS) model back in 2025 — but it wasn’t accessible to most of us, as it required getting onto a dedicated enterprise allowlist. Then, in September 2026, Google brought <a href="https://aistudio.google.com/docs/voice-replication?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Voice Replication</a> and <a href="https://aistudio.google.com/docs/voice-design?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Voice Design</a> to the masses with <strong>Gemini 3.8 Flash TTS</strong> (gemini-3.8-flash-tts). Now <strong>anyone can clone their own voice</strong>, create custom vocal personas from a prompt, and add rich style guidance and inline vocal expressions like whispering, laughing, sighing, and pausing.</p><p>Today I decided to give this a try. And because I’ve been doing quite a few blogs and talks about <strong>Harness Engineering</strong> lately, I decided to put this into practice. So, the goal today was to build a clean voice cloning application from scratch, using <strong>Google Antigravity</strong>, and leveraging my existing coding harness. The application will:</p><ul><li>Quickly and easily clone a supplied voice, handling audio normalisation and mandatory consent checks</li><li>Persist and manage my voice clones in Google Cloud</li><li>Provide a simple interface for generating expressive monologue and multi-speaker dialogue from text in my voice</li></ul><p>So, grab your coffee and your cat and let’s see how to go from nothing to a fully tested, documented and public-repo-ready application in one hour!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*soqcFuTun8ae3zE6.png" /></figure><h3>Why Build a Whole Application Around a Couple of API Calls?</h3><p>At first glance, calling the Gemini API’s /v1beta/voices endpoint is pretty simple. You send two audio clips to client.voices.create() — a reference recording of your voice and a recording of you reciting a mandatory consent statement — and you get back a voice identifier (voice_...) that you can pass to client.interactions.create() using gemini-3.8-flash-tts.</p><blockquote>“Wait — if it’s that simple, why build a whole CLI and storage layer around it?”<em> </em>“Partly for convenience… and mostly to give my agent harness a proper workout!”</blockquote><p>Let’s be honest: if you only want a one-off voice clone, a 20-line Python script is probably enough. (Interestingly, although the docs mention an interactive Voice Replication experience in Google AI Studio, that option wasn’t actually showing up in the AI Studio UI for me in the UK.) My main motivation for building a full CLI around all this was to test how quickly my <strong>Google Antigravity</strong> harness could turn a loose idea into a well-engineered, tested tool.</p><p>That said, wrapping the raw API in a proper CLI also solves a few everyday annoyances:</p><ul><li><strong>Automatic audio conversion</strong>: For best results, Google recommends <strong>24kHz, mono, 16-bit PCM WAV</strong> audio and a <strong>10 to 30 second</strong> reference clip. Meanwhile, your phone or desktop recorder probably spits out 48kHz stereo .m4a or .mp3 files. Having the CLI automatically convert and check your clips via ffmpeg saves manual conversion steps.</li><li><strong>Consent sanity checking</strong>: To prevent unauthorised voice cloning, you must record a verbatim consent phrase (&quot;I am the owner of this voice and I consent to Google using this voice to create a synthetic voice model.&quot;). Sure, if you upload the wrong file or fluff a word, the Voices endpoint will just reject it — hardly the end of the world! — but checking it first with gemini-3.5-transcribe tells you <em>exactly</em> which words didn&#39;t match, and gave us another fun Gemini API integration to play with.</li><li><strong>Voice persistence and TTL expiry</strong>: When you create a stateful voice (store=True), Google automatically persists it inside the Google Cloud project tied to your GEMINI_API_KEY with a <strong>1-year retention TTL</strong> (or <strong>7 days</strong> if you use a stateless voicekey_... token with store=False). Adding an optional Google Cloud Storage backup alongside your local voices.json catalogue means archiving your master WAV clips — or refreshing an expired voice a year from now from another machine — won&#39;t be a PITA.</li><li><strong>Expressive multi-speaker scripting</strong>: Gemini 3.8 Flash TTS supports turn-level style instructions (speech_metadata.style — e.g. &quot;whispering&quot; or &quot;excited&quot;) and inline vocal tags like &lt;laugh&gt;, &lt;sigh&gt;, and &lt;short pause&gt;. Typing raw JSON payloads by hand gets old fast; feeding a clean Markdown monologue or YAML dialogue script into a CLI is much nicer.</li></ul><p>Best of all, because the entire workflow is wrapped behind simple make targets (including make design if you also want to generate brand-new synthetic voices from a text prompt), you can clone <a href="https://github.com/derailed-dash/voice-replication">derailed-dash/voice-replication</a> and use it out of the box without touching a line of Python.</p><h3>Starting with a Reusable Harness: GEMINI.md, Skills, MCPs, and Plugins</h3><p>Before I even typed my first prompt, Antigravity already knew <em>how</em> I build software. Why? Because of my reusable ~/.gemini harness.</p><p>As I covered in <a href="https://dev.to/gde/from-vibing-chaos-to-reliable-agents-the-what-why-and-how-of-harness-engineering-4ba0">From Vibing Chaos to Reliable Agents: The What, Why, and How of Harness Engineering</a>, separating <strong>Context (<em>HOW</em>)</strong> from <strong>Specification (<em>WHAT</em>)</strong> is the foundation for getting consistent, production-grade output from an AI coding agent. Instead of repeating my coding standards in every prompt, my global harness combines a GEMINI.md engineering constitution with plugins, skills, and MCP servers across every repository I open:</p><ul><li><strong>Modern Python &amp; </strong><strong>uv</strong>: Enforces Python 3.13, uv environment management, and modern type hinting.</li><li><strong>Latest GA Gemini Models via the Interactions API</strong>: Mandates current models and never falls back to legacy generateContent endpoints.</li><li><strong>Plugins, On-Demand Skills, &amp; Live MCP Servers</strong>: Pairs bundled plugins (gemini-api, google-cloud-developer) and live documentation MCPs (gemini-api-docs, developer-knowledge) with curated skills.</li><li><strong>Self-Documenting </strong><strong>Makefile &amp; </strong><strong>.env.template Hygiene</strong>: Every repo gets a Makefile, a sanitised .env.template, and rules to ensure nothing sensitive ever gets pushed to GitHub.</li><li><strong>Spec-Driven Development, TDD, &amp; Self-Correcting Loops</strong>: Every non-trivial feature starts with SPEC.md, TODO.md, and LOOPS.md, ensuring our harness verifies each iteration both deterministically and semantically.</li></ul><h3>Interactive Discovery: Using /grill-me Before Writing a Line of Code</h3><p>I kicked off the session with a single prompt using the /grill-me slash command:</p><pre>/grill-me I want to build a project for managing my personal voice replication,<br>using gemini flash tts, and the guidance here:<br>https://aistudio.google.com/docs/voice-replication.<br>I want to persist my voice clone in my GCP project.<br>We will use Python. I can provide audio clips, and the mandatory recorded statement.</pre><p>Instead of immediately hallucinating half-baked Python files, Antigravity followed my harness protocol:</p><ol><li><strong>Activated Relevant Skills</strong>: It read gemini-api-dev, google-cloud-storage-bucket-architect, google-cloud-storage-basics, and retrieving-developer-knowledge.</li><li><strong>Queried Live Upstream Docs via MCP</strong>: It pulled the current documentation for <a href="https://aistudio.google.com/docs/voice-replication?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Voice Replication</a>, <a href="https://aistudio.google.com/docs/speech-generation?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Speech Generation</a>, <a href="https://aistudio.google.com/docs/voice-design?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Voice Design</a>, and <a href="https://aistudio.google.com/docs/transcribe?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Audio Transcription</a> from the Gemini API Docs MCP server, and checked <a href="https://docs.cloud.google.com/storage/docs/storage-classes?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Cloud Storage Classes</a>, <a href="https://docs.cloud.google.com/storage/docs/autoclass?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Autoclass</a>, and <a href="https://cloud.google.com/storage/pricing?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Pricing</a> via the Developer Knowledge MCP server.</li><li><strong>Conducted a Structured Design Interview</strong>: Using multiple-choice questions, it interviewed me across several architectural trade-offs to lock in concrete design decisions.</li></ol><p>The resulting solution — fully generated by Gemini Argon inside Antigravity — looks like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*L7op92V9m2DcDq-U.png" /></figure><h3>Locking the Blueprint: SPEC.md, TODO.md, and LOOPS.md</h3><p>Once the /grill-me interview wrapped up, Antigravity activated my spec-driven-development, planning-and-task-breakdown, agent-aware-cli, and <a href="https://github.com/Forward-Future/loopy/tree/main/skills/loopy">loopy</a> skills to generate the project&#39;s specification artefacts on disk <em>before</em> writing any code:</p><ul><li><strong>SPEC.md</strong>: Captured the system architecture, Gemini 3.8 TTS &amp; Gemini 3.5 Transcribe API contracts, the ffmpeg 24kHz mono WAV pipeline, the hybrid GCS Coldline/Standard persistence design, the CLI command table, explicit operational boundaries (Always Do / Ask First / Never Do), and the complete two-tier verification rubric.</li><li><strong>tasks/plan.md &amp; </strong><strong>TODO.md</strong>: Decomposed the build into eight independently testable vertical slices with clear verification commands and phase checkpoints.</li><li><strong>LOOPS.md</strong>: Saved the self-correcting <strong>TDD Vertical Slice Loop</strong> prompt and its binary acceptance rubric, so the loop can be re-run at any time.</li></ul><p>As I wrote in <a href="https://dev.to/gde/spec-driven-development-for-the-win-disposable-code-living-blueprints-and-building-like-a-boss-2bnn">Spec-Driven Development For-the-Win: Disposable Code, Living Blueprints, and Building Like a Boss</a>, <strong>code is disposable; the specification is what matters</strong>. Pausing at a hard-stop approval gate to review SPEC.md and LOOPS.md took two minutes — and meant that when I typed y to start building, the agent had zero ambiguity left to guess about.</p><h3>Executing with Loopy + TDD: Red, Green, Refactor</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Jkw87h7j9N55teLv.png" /></figure><p>With the spec approved, Antigravity activated <a href="https://github.com/Forward-Future/loopy/tree/main/skills/loopy">loopy</a> and test-driven-development and executed each vertical slice in TODO.md following our loop cycle: <strong>Observe -&gt; Choose -&gt; Act -&gt; Verify -&gt; Record</strong>.</p><p>If you read <a href="https://dev.to/gde/loop-engineering-in-the-harness-autonomous-self-correcting-systems-part-2-155e">Loop Engineering in the Harness: Autonomous, Self-Correcting Systems</a>, you will recognise the mechanics immediately:</p><ul><li><strong>Red Phase First</strong>: For every slice, Antigravity wrote the failing pytest test file in tests/ <em>first</em> and ran uv run pytest to prove it failed for the right reason.</li><li><strong>Iteration Discipline (The One-Change Rule)</strong>: Each pass implemented a single semantic slice (or fixed one behavioural bug) before re-verifying. If a check failed, the agent knew <em>exactly</em> which change caused the issue — whereas purely mechanical linting or formatting fixes could be batched together.</li><li><strong>Fast, Offline Testing</strong>: Audio tests run against real ffmpeg using tiny generated test clips, while Gemini API and Cloud Storage calls use injected fakes — so the entire 38-test suite runs offline in about 3 seconds and with 0 API cost.</li></ul><h3>Two-Tier Verification: Mechanical Short-Circuiting + Stateless Semantic Critic</h3><p>Left to its own devices, an agent will happily write some questionable code, quietly violate your architectural rules, and declare victory. Why? Because the same agent that wrote the code is <strong>grading its own homework inside a polluted context window</strong>.</p><p>To fix this — and avoid burning tokens on AI code reviews when basic tests or linters are still failing — my harness checks the work in two stages:</p><ol><li>Run the fast, free, deterministic tests first.</li><li>Only bring in a separate “Critic” agent once the basics pass.</li></ol><h4>Tier 1: The Deterministic Mechanical Gate</h4><p>After every slice, Antigravity runs make check, which executes four deterministic tools in sequence:</p><pre>uv run codespell -s          # catch spelling mistakes in code and docs<br>uv run ruff check --fix .    # lint and auto-fix formatting issues<br>uv run pyright               # static type checking<br>uv run pytest -m &quot;not live&quot;  # run fast offline unit tests (skip live cloud calls in the inner loop)</pre><p>If any Tier 1 check fails — a single typo caught by codespell, an unused import in ruff, a type mismatch in pyright, or tests failing — the loop <strong>short-circuits immediately</strong> and fixes the mechanical defect before any LLM review is attempted.</p><h4>Tier 2: The Stateless Semantic Critic</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*SL3Pk4GE2VJEVzn1.png" /></figure><p>Once all slices pass Tier 1, the loop spawns a <strong>fresh, stateless </strong><strong>research subagent</strong> with zero builder conversation history. That critic reads the codebase cold and grades it against the numbered binary assertions codified in SPEC.md and LOOPS.md:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fsiteproxy.ruqli.workers.dev%3A443%2Fhttps%2Fdatawrapper.dwcdn.net%2FULQ1J%2F1%2F&amp;display_name=Datawrapper&amp;url=https%3A%2F%2Fsiteproxy.ruqli.workers.dev%3A443%2Fhttps%2Fdatawrapper.dwcdn.net%2FULQ1J%2F1%2F&amp;image=https%3A%2F%2Fsiteproxy.ruqli.workers.dev%3A443%2Fhttps%2Fdatawrapper.dwcdn.net%2FULQ1J%2Fplain-s.png%3Fv%3D1&amp;type=text%2Fhtml&amp;schema=dwcdn" width="600" height="624" frameborder="0" scrolling="no"><a href="https://medium.com/media/21a247add24968f7942fc02b94ad47ee/href">https://medium.com/media/21a247add24968f7942fc02b94ad47ee/href</a></iframe><p>When the stateless Critic finishes its audit and returns <strong>7 / 7 PASS (100%)</strong>, loopy writes the final evidence receipt to .cache/loops/last_receipt.md and prints the run summary right in chat:</p><pre>Loopy — TDD Vertical Slice Loop<br>========================================<br>Stop reason : target_met (7 vertical slices completed)<br><br>Iter  Before  After   Delta   Kept  Summary<br>1     0/11    11/11   +11     yes   Config &amp; Pydantic models (config.py, models.py, test_models.py)<br>2     11/17   17/17   +6      yes   FFmpeg 24kHz mono WAV normaliser &amp; validator (audio.py, test_audio.py)<br>3     17/22   22/22   +5      yes   Consent phrase registry &amp; gemini-3.5-transcribe gate (consent.py, test_consent.py)<br>4     22/28   28/28   +6      yes   Gemini Voices CRUD &amp; TTS synthesis client (client.py, test_client.py)<br>5     28/32   32/32   +4      yes   Local catalogue &amp; GCS Coldline/Standard sync (storage.py, test_storage.py)<br>6     32/38   38/38   +6      yes   Agent-Aware Rich/Typer CLI &amp; examples (cli.py, test_cli.py, examples/)<br>7     6/7     7/7     +1      yes   Core docs, PNG architecture diagram &amp; Tier 2 Stateless Semantic Critic audit</pre><p>Well, this is pretty cool to see. Especially since I didn’t write any of the rubrics or looping instructions by hand for this project. And this output appeared literally minutes after kicking off the build!</p><h3>Developer Experience: A Self-Documenting Makefile and Agent-Aware CLI</h3><p>Because my GEMINI.md harness pairs the agent-aware-cli skill with a self-documenting Makefile that pulls its configuration dynamically from .env.template and .env, running the application is effortless whether you are a human in a terminal or an AI agent invoking tools.</p><p>Typing make (or make help) prints every verification gate and application workflow alongside the active .env configuration:</p><pre>Usage:<br>  make &lt;target&gt; [VARIABLE=value ...]<br><br>Development &amp; Verification Gates<br>  install             Create/sync virtual environment and install dependencies via uv<br>  lint                Run deterministic static checks (codespell, ruff, pyright)<br>  fix                 Auto-fix formatting and lint issues via ruff and check spelling<br>  test                Run offline unit test suite with pytest and coverage<br>  check               Run full Tier 1 verification gate (codespell + ruff + pyright + pytest)<br>Audio Pre-Flight &amp; Voice Lifecycle<br>  prep-audio          Normalise reference &amp; consent clips to 24kHz mono WAV and verify consent<br>  create              Create a stateful replicated voice (store=True, 1-year TTL)<br>  design              Create a custom synthetic voice from a text prompt<br>  list / inspect      List or inspect voices in the local catalogue (data/voices.json)<br>  refresh / delete    Renew a voice&#39;s 1-year TTL from master WAVs or delete a voice<br>Speech &amp; Dialogue Synthesis<br>  speak               Synthesise single-speaker audio (FILE=monologue.md)<br>  dialogue            Synthesise multi-speaker dialogue from YAML (SCRIPT=dialogue.yaml)<br>Google Cloud Storage (GCS) Backup &amp; Recovery<br>  gcs-preflight       Display read-only gcloud pre-flight verification commands<br>  gcs-init            Provision regional GCS backup bucket<br>  sync-push / pull    Sync voices.json and audio WAVs with GCS</pre><h3>Solution Code Snippets</h3><p>Let’s take a quick peek under the hood at some of the code behind make create.</p><h4>1. Verifying the Mandatory Consent Recording (consent.py)</h4><p>The function verify_consent_audio() sends the normalised consent WAV to gemini-3.5-transcribe via the Gemini Interactions API (client.interactions.create) to verify that the spoken words match Google&#39;s mandatory consent statement before we ever call the Voices endpoint:</p><pre>def verify_consent_audio(<br>    audio_path: Path,<br>    *,<br>    client: Any,<br>    language_code: str = &quot;en-GB&quot;,<br>    model: str = &quot;gemini-3.5-transcribe&quot;,<br>    threshold: float = DEFAULT_SIMILARITY_THRESHOLD,<br>) -&gt; ConsentCheckResult:<br>    &quot;&quot;&quot;Transcribe `audio_path` with `gemini-3.5-transcribe` and verify consent wording.&quot;&quot;&quot;<br>    expected_phrase = get_consent_phrase(language_code)<br>    audio_b64 = base64.b64encode(audio_path.read_bytes()).decode(&quot;utf-8&quot;)<br><br>    bcp47_hint = &quot;en-GB&quot; if language_code == &quot;en&quot; else language_code<br>    interaction = client.interactions.create(<br>        model=model,<br>        input=[<br>            {<br>                &quot;type&quot;: &quot;audio&quot;,<br>                &quot;data&quot;: audio_b64,<br>                &quot;mime_type&quot;: &quot;audio/wav&quot;,<br>            }<br>        ],<br>        generation_config={<br>            &quot;transcription_config&quot;: {<br>                &quot;language_codes&quot;: [bcp47_hint],<br>            }<br>        },<br>    )<br>    transcript = (getattr(interaction, &quot;output_text&quot;, None) or &quot;&quot;).strip()<br>    similarity = compute_phrase_similarity(expected_phrase, transcript)<br>    return ConsentCheckResult(<br>        passed=similarity &gt;= threshold,<br>        language_code=language_code,<br>        expected_phrase=expected_phrase,<br>        transcript=transcript,<br>        similarity=round(similarity, 4),<br>        audio_path=audio_path,<br>    )</pre><h4>2. Creating &amp; Persisting the Cloned Voice (client.py)</h4><p>Before looking at our wrapper method, it’s worth pausing to see <strong>just how trivial the underlying Gemini API call actually is</strong>. Once you have your two WAV files (source.wav and consent.wav), creating a persistent voice clone on Google&#39;s servers takes barely a dozen lines of Python:</p><pre>from google import genai<br><br>client = genai.Client(http_options={&quot;api_version&quot;: &quot;v1beta&quot;})<br><br>response = client.voices.create(<br>    store=True,<br>    voice={<br>        &quot;model&quot;: &quot;gemini-3.8-flash-tts&quot;,<br>        &quot;type&quot;: &quot;replicated&quot;,<br>        &quot;display_name&quot;: &quot;Dazbo Primary&quot;,<br>        &quot;replicated&quot;: {<br>            &quot;source_audio&quot;: {&quot;mime_type&quot;: &quot;audio/wav&quot;, &quot;data&quot;: source_b64},<br>            &quot;consent_audio&quot;: {&quot;mime_type&quot;: &quot;audio/wav&quot;, &quot;data&quot;: consent_b64},<br>        },<br>    },<br>)<br>print(response.id)  # e.g. voice_... (stored in your GCP project for 1 year!)</pre><p>That’s literally it! Passing <strong>store=True</strong> tells the Gemini API to persist the cloned voice inside the Google Cloud project associated with your GEMINI_API_KEY for a full year (and automatically makes it show up in Google AI Studio) — no database or storage bucket required.</p><p>In client.py, our create_replicated_voice() method simply wraps that exact self.client.voices.create() call and returns a typed VoiceRecord model (which calculates the 1-year TTL expiry countdown and gets saved to our local data/voices.json catalogue so we can refer to the voice by a friendly alias like dazbo instead of pasting raw voice_... strings everywhere):</p><pre>def create_replicated_voice(<br>    self,<br>    *,<br>    alias: str,<br>    display_name: str,<br>    source_wav_path: Path,<br>    consent_wav_path: Path,<br>    store: bool = True,<br>    model: str = &quot;gemini-3.8-flash-tts&quot;,<br>    default_style: str | None = None,<br>    language_code: str = &quot;en-GB&quot;,<br>) -&gt; VoiceRecord:<br>    &quot;&quot;&quot;Create a stateful (`store=True`) or stateless (`store=False`) replicated voice.&quot;&quot;&quot;<br>    source_b64 = base64.b64encode(source_wav_path.read_bytes()).decode(&quot;utf-8&quot;)<br>    consent_b64 = base64.b64encode(consent_wav_path.read_bytes()).decode(&quot;utf-8&quot;)<br><br>    voice_payload: dict[str, Any] = {<br>        &quot;model&quot;: model,<br>        &quot;type&quot;: &quot;replicated&quot;,<br>        &quot;replicated&quot;: {<br>            &quot;source_audio&quot;: {&quot;mime_type&quot;: &quot;audio/wav&quot;, &quot;data&quot;: source_b64},<br>            &quot;consent_audio&quot;: {&quot;mime_type&quot;: &quot;audio/wav&quot;, &quot;data&quot;: consent_b64},<br>        },<br>    }<br>    if store:<br>        voice_payload[&quot;display_name&quot;] = display_name<br><br>    response = self.client.voices.create(store=store, voice=voice_payload)<br>    return VoiceRecord(<br>        alias=alias,<br>        display_name=display_name,<br>        voice_type=VoiceType.REPLICATED,<br>        storage_mode=StorageMode.STATEFUL if store else StorageMode.STATELESS,<br>        voice_id=getattr(response, &quot;id&quot;, None),<br>        voice_key=getattr(response, &quot;key&quot;, None),<br>        model=model,<br>        default_style=default_style,<br>        language_code=language_code,<br>        source_audio_path=str(source_wav_path),<br>        consent_audio_path=str(consent_wav_path),<br>    )</pre><h3>Taking It for a Spin!</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*wIod3_UKoaFv8AAv.png" /></figure><h3>A Quick Note on Dependencies (No GCS Bucket Required!)</h3><p>If you look at .env.template or Makefile in <a href="https://github.com/derailed-dash/voice-replication">derailed-dash/voice-replication</a>, you&#39;ll notice variables like GCP_PROJECT_ID and GCS_BUCKET_NAME alongside targets like make gcs-init and make sync-push.</p><p>If you just want to do some quick voice cloning and speech generation, <strong>you can ignore the Google Cloud Storage bucket and </strong><strong>gcloud setup completely</strong>:</p><ul><li><strong>Only the </strong><strong>GEMINI_API_KEY is required</strong>: To run make prep-audio, make create, make speak, and make dialogue, all you need locally are uv, ffmpeg, and a <strong>GEMINI_API_KEY</strong> in your .env file (created in <a href="https://aistudio.google.com/apikey?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Google AI Studio</a>).</li><li><strong>Paid Quota Tier 1</strong>: One small catch to be aware of — Gemini Voice Replication (gemini-3.8-flash-tts) requires <strong>Paid Quota Tier 1 or higher</strong>. That simply means the Google Cloud project behind your AI Studio API key must have a Cloud Billing account linked.</li><li><strong>Cloud Storage is 100% optional</strong>: Because make create passes store=True, the Gemini API <em>already</em> persists your cloned voice server-side in your Google Cloud project, while voice-rep saves your alias and voice_... ID locally in data/voices.json. The google-cloud-storage client is only ever touched if you explicitly run make gcs-init or make sync-push to back up your master .wav files to Coldline storage.</li></ul><h4>1: Cloning My Voice</h4><p>To create a persistent cloned voice via the Gemini API’s /v1beta/voices endpoint (client.voices.create()), we first need two short recordings:</p><ul><li><strong>A Reference Audio Clip (10 to 30 seconds)</strong>: Around 20 to 25 seconds of natural, expressive conversational speech so Gemini can capture your natural cadence, pitch variation, and warmth.</li><li><strong>The Mandatory Consent Clip</strong>: A clear recording of the same voice reciting this exact consent phrase: &quot;I am the owner of this voice and I consent to Google using this voice to create a synthetic voice model.&quot;</li></ul><p>I recorded these short clips using <a href="https://www.audacityteam.org/">Audacity</a> (the free, open-source audio editor) and exported them as standard 16-bit WAV files.</p><p>Now I ran these commands to prepare, create, and back up my dazbo voice:</p><pre>make prep-audio ALIAS=dazbo SOURCE=data/raw/dazbo_ref.wav CONSENT=data/raw/dazbo_consent.wav<br>make create ALIAS=dazbo DISPLAY_NAME=&quot;Dazbo Primary&quot;<br>make sync-push # optional: backs up master WAVs and voices.json to GCS</pre><ul><li>make prep-audio resamples each recording as required and sends the consent clip to gemini-3.5-transcribe, verifying a 100% word match against the mandatory statement.</li><li>make create sends the reference and consent WAVs to client.voices.create() (store=True), which creates the persistent cloned voice in our Google Cloud project and saves the returned voice_... ID to our local voices.json catalogue under the alias dazbo.</li><li>make sync-push <em>(optional)</em> backs up the local voices.json catalogue and normalised WAV files to Google Cloud Storage if you&#39;ve configured a backup bucket.</li></ul><h4>2: Text-to-Speech</h4><p>To prove the voice works as expected, I tested it against a short Markdown monologue script (examples/monologue.md):</p><blockquote>“Well howdy. You’re not listening to Dazbo. You’re listening to his voice clone, created using gemini-3.8-flash-tts. In a few minutes, you can create your own voice clone. Enjoy!”</blockquote><pre>make speak ALIAS=dazbo FILE=examples/monologue.md OUT=output/monologue.mp3</pre><p>Within a few seconds, output/monologue.mp3 was sitting in output/. Here is the full flow in action — running the make commands and playing back the synthesised voice clone:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fsiteproxy.ruqli.workers.dev%3A443%2Fhttps%2Fwww.youtube.com%2Fembed%2F2C-UQWglvxw%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fsiteproxy.ruqli.workers.dev%3A443%2Fhttps%2Fwww.youtube.com%2Fwatch%3Fv%3D2C-UQWglvxw&amp;image=https%3A%2F%2Fsiteproxy.ruqli.workers.dev%3A443%2Fhttps%2Fi.ytimg.com%2Fvi%2F2C-UQWglvxw%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/8b89796d16e7ab65d3a4e4aa511fb663/href">https://medium.com/media/8b89796d16e7ab65d3a4e4aa511fb663/href</a></iframe><p><em>THIS IS SO COOL!</em></p><h4>3: Cloning a Second Voice &amp; Demoing a Two-Person Story — The Great Pie Heist</h4><p>Once the single-voice demo was proven, my wife Kate wanted to try it too. We recorded her reference and consent clips, ran the exact same make prep-audio and make create ALIAS=kate DISPLAY_NAME=&quot;Kate Primary&quot; commands, and decided to test building a two-person conversation with our cloned voices in a single Interactions API call.</p><p>Remember at the start of this article when I told you to grab your coffee and your cat? Well, we have two cats: <strong>Leia</strong> (aka <strong>Laylay</strong>) and <strong>Mycroft</strong> (aka <strong>Mikey</strong>).</p><p>I asked Antigravity to create a 2-minute dialogue script about our kitties, in the required yaml format. It created this:</p><pre>model: gemini-3.8-flash-tts<br>speakers:<br>  Dazbo: dazbo<br>  Kate: kate<br>turns:<br>  - speaker: Dazbo<br>    text: &quot;Honestly, Kate, I think we need to stage a full-scale intervention. I went into the kitchen five minutes ago to grab a coffee, and the crime scene on the counter was just tragic.&quot;<br>    style: &quot;deadpan, amused British storyteller&quot;<br>  - speaker: Kate<br>    text: &quot;Oh no. Don&#39;t tell me. The steak and ale pie?&quot;<br>    style: &quot;knowing, slightly exasperated but amused&quot;<br>  - speaker: Dazbo<br>    text: &quot;Reduced to a mere crust, my love. And right next to the tin — not even attempting to flee the scene, mind you — was Laylay. Fast asleep, with flaky pastry crumbs dusted all over her whiskers.&quot;<br>    style: &quot;expressive, conversational, slightly dramatic&quot;<br><br># ... you get the idea! (Full 14-turn script in the repo)</pre><p>And then I ran this single command:</p><pre>make dialogue SCRIPT=data/scripts/laylay_and_mikey.yaml OUT=output/laylay_and_mikey.mp3</pre><p>About ten seconds later, the mp3 file popped out. Check it out!</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fsiteproxy.ruqli.workers.dev%3A443%2Fhttps%2Fwww.youtube.com%2Fembed%2FzYMSKLJCLKU%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fsiteproxy.ruqli.workers.dev%3A443%2Fhttps%2Fwww.youtube.com%2Fwatch%3Fv%3DzYMSKLJCLKU&amp;image=https%3A%2F%2Fsiteproxy.ruqli.workers.dev%3A443%2Fhttps%2Fi.ytimg.com%2Fvi%2FzYMSKLJCLKU%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/c2ef04d1761be43951d00788828f9966/href">https://medium.com/media/c2ef04d1761be43951d00788828f9966/href</a></iframe><h4>4: Using Our Cloned Voices Directly in Google AI Studio</h4><p>That immediately begged the next question: <em>“Now that we’ve created these voices, can I use them in the Google AI Studio web UI?”</em></p><p>The answer is <strong>yes</strong>! Remember earlier when I mentioned that the option to <em>create</em> a cloned voice wasn’t showing up in the Google AI Studio UI for me in the UK? Well, because make create registers voices via the API with <strong>store=True</strong>, those voices are saved directly into the Google Cloud project&#39;s <strong>Gemini Voices catalogue</strong> — which means they immediately show up ready to use inside Google AI Studio anyway!</p><p>It’s as easy as 1–2–3!</p><ol><li>Open <a href="https://aistudio.google.com/generate-speech?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">https://aistudio.google.com/generate-speech</a> and <strong>select the GCP project</strong> where the voices were created.</li><li>Click <strong>+ Create new dialog</strong>.</li><li><strong>Pick Your Cloned Voices under Speaker Settings</strong>: In the right-hand <strong>Run settings</strong> panel under <strong>Speaker settings</strong> (or by clicking the speaker pill on any speech block), open the voice selector. Right at the very top under <strong>Custom voices</strong> — above all the built-in voices (Bodi, Lumi, Sola, Koda) — sit <strong>Dazbo Primary</strong> and <strong>Kate Primary</strong>!</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*9Hexu4LkthFb5rpl.jpg" /></figure><p>That means we get the best of both worlds: a low-effort, type-and-go local CLI (make speak, make dialogue, etc.) for scripted stories, plus Google&#39;s native web experience in Google AI Studio whenever we just want to type a quick prompt and hit <strong>Run</strong>.</p><h3>Wrapping Up: Come for the Voice Clones, Stay for the Harness</h3><p>Building <a href="https://github.com/derailed-dash/voice-replication">voice-replication</a> in a single sitting proved two things for me.</p><p>First, <strong>every layer of the Google Antigravity harness paid for itself</strong>:</p><ul><li><strong>Global Context (</strong><strong>~/.gemini/GEMINI.md + Skills + MCPs)</strong> supplied the reusable engineering standards and live Google documentation before I even typed a prompt.</li><li><strong>Interactive Discovery (</strong><strong>/grill-me + </strong><strong>SPEC.md)</strong> resolved the audio normalisation, consent transcription, and Cloud Storage class trade-offs before a line of code was written.</li><li><strong>Self-Correcting Execution (</strong><strong>LOOPS.md + TDD + Tier 1 &amp; Tier 2 Gates)</strong> drove every vertical slice through Red -&gt; Green -&gt; Refactor, short-circuiting mechanical errors at $0 cost before a stateless Critic verified the rubric.</li></ul><p>Second, <strong>you don’t need to write any code to try this yourself</strong>. If you want to clone your own voice, script multi-speaker podcasts, or unlock your custom voices inside Google AI Studio:</p><ol><li><strong>Clone the repo</strong>: Grab <a href="https://github.com/derailed-dash/voice-replication">derailed-dash/voice-replication</a> on GitHub (and drop it a star ⭐ while you&#39;re there!).</li><li><strong>Record two short clips</strong>: Record 20 seconds of natural speech and the mandatory consent sentence.</li><li><strong>Run two commands</strong>: Run make prep-audio and make create — and within a couple of minutes, your voice clone will be ready for both terminal scripting (make speak, make dialogue) and the Google AI Studio UI!</li></ol><p>Give it a spin, and let me know in the comments what you end up building (or which family members and pets you end up scripting dialogues about)!</p><h3>Before You Go</h3><ul><li><strong>Please share</strong> 📢 this with anyone that you think will be interested. It might help them, and it really helps me!</li><li>Please <strong>give me loads of claps</strong>! 👏 (Just hold down the clap button.)</li><li>Please <strong>leave a comment</strong> 💬. Interaction is good!</li><li><strong>Add a star</strong> ⭐ on my repos!</li><li><strong>Follow</strong> 👉 and <strong>subscribe</strong> 🔔, so you don’t miss my content.</li></ul><h3>References and Useful Links</h3><ul><li><a href="https://github.com/derailed-dash/voice-replication">Source Code Repository — </a><a href="https://github.com/derailed-dash/voice-replication">derailed-dash/voice-replication</a></li><li><a href="https://aistudio.google.com/docs/voice-replication?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Gemini API — Voice Replication Guide</a></li><li><a href="https://aistudio.google.com/docs/speech-generation?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Gemini API — Speech Generation (TTS) Guide</a></li><li><a href="https://aistudio.google.com/docs/voice-design?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Gemini API — Voice Design Guide</a></li><li><a href="https://aistudio.google.com/docs/transcribe?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Gemini API — Audio Transcription (</a><a href="https://aistudio.google.com/docs/transcribe?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">gemini-3.5-transcribe)</a></li><li><a href="https://docs.cloud.google.com/storage/docs/storage-classes?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Google Cloud Storage — Storage Classes</a></li><li><a href="https://signals.forwardfuture.com/loop-library/">The Loop Library</a> &amp; <a href="https://github.com/Forward-Future/loopy/tree/main/skills/loopy">loopy Skill</a></li><li><a href="https://dev.to/gde/from-vibing-chaos-to-reliable-agents-the-what-why-and-how-of-harness-engineering-4ba0">From Vibing Chaos to Reliable Agents: The What, Why, and How of Harness Engineering (Part 1)</a></li><li><a href="https://dev.to/gde/loop-engineering-in-the-harness-autonomous-self-correcting-systems-part-2-155e">Loop Engineering in the Harness: Autonomous, Self-Correcting Systems (Part 2)</a></li><li><a href="https://dev.to/gde/spec-driven-development-for-the-win-disposable-code-living-blueprints-and-building-like-a-boss-2bnn">Spec-Driven Development For-the-Win: Disposable Code, Living Blueprints, and Building Like a Boss (Part 3)</a></li><li><a href="https://dev.to/gde/skills-sprawl-when-too-much-of-a-good-thing-confuses-your-ai-agent-4nij">Skills Sprawl: When Too Much of a Good Thing Confuses Your AI Agent</a></li><li><a href="https://dev.to/gde/meet-the-antigravity-extension-gemini-code-assist-on-steroids-in-whatever-ide-you-love-4p1n">Meet the Antigravity Extension: Gemini Code Assist on Steroids, in Whatever IDE You Love</a></li><li><a href="https://darrenlester.net">Dazbo’s Portfolio</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7ffa3e426f0f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/google-cloud/how-i-made-a-voice-clone-in-minutes-using-google-antigravity-gemini-3-8-flash-tts-and-my-harness-7ffa3e426f0f">How I Made a Voice Clone in Minutes, Using Google Antigravity, Gemini 3.8 Flash TTS and My Harness</a> was originally published in <a href="https://medium.com/google-cloud">Google Cloud - Community</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Dazbo’s Technical Writing & Engineering Catalogue: Agentic AI, Google Cloud, Architecture, and…]]></title>
            <link>https://medium.com/@dazbo/dazbos-technical-writing-engineering-catalogue-agentic-ai-google-cloud-architecture-and-cd21558dde32?source=rss-58a0bd8ba726------2</link>
            <guid isPermaLink="false">https://medium.com/p/cd21558dde32</guid>
            <category><![CDATA[gemini]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[google-antigravity]]></category>
            <category><![CDATA[google-cloud]]></category>
            <category><![CDATA[cloud-architecture]]></category>
            <dc:creator><![CDATA[Dazbo (Darren Lester)]]></dc:creator>
            <pubDate>Sun, 08 Nov 2026 21:44:31 GMT</pubDate>
            <atom:updated>2026-11-08T22:05:37.477Z</atom:updated>
            <content:encoded><![CDATA[<h3>Dazbo’s Technical Writing &amp; Engineering Catalogue: Agentic AI, Google Cloud, Architecture, and Python</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Ws67GlP9CdzXAnJ6.png" /></figure><p>Hey friends! Welcome to my <strong>Start Here</strong> directory for everything I write and build as <strong>Dazbo (Darren Lester)</strong> — Google Chief Technologist, Google Developer Expert (GDE), Google Cloud Ambassador, and Enterprise Cloud Architect. Here you’ll find articles relating to Google Cloud and AI, architecture, and software engineering.</p><p>Whether you’re after a quick five-minute explainer to demystify a new concept, a pragmatic architectural blueprint to help you navigate real-world trade-offs, or a deep-dive technical walkthrough where we build and deploy a production system from scratch, you’ll find it organised by theme below. So grab a coffee, pick a topic, and dive in!</p><blockquote><strong><em>Interactive Portfolio &amp; AI Assistant:</em></strong><em> If you want to search across all of my articles, open-source repositories, and architectural demos, then you can ask my AI agent. Head over to </em><a href="https://darrenlester.net"><strong><em>darrenlester.net</em></strong></a><em> (and the full searchable </em><a href="https://darrenlester.net/blogs"><strong><em>Blogs Directory</em></strong></a><em>).</em></blockquote><h3>1. Agentic Engineering, Harness Engineering &amp; Spec-Driven Development</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*jvpVjBxkEyXEjoWv.png" /></figure><p>Moving from fragile “vibe coding” to deterministic, self-correcting engineering harnesses that produce production-grade software:</p><ul><li><a href="https://medium.com/google-cloud/spec-driven-development-for-the-win-disposable-code-living-blueprints-and-building-like-a-boss-7fe52e7d2d49"><strong>Spec-Driven Development For-the-Win: Disposable Code, Living Blueprints, and Building Like a Boss</strong></a> — Why specifications are the true source of truth in the agentic era, and how to decompose requirements into living blueprints before writing a line of code.</li><li><a href="https://medium.com/google-cloud/loop-engineering-in-the-harness-autonomous-self-correcting-systems-95d3ac8f64cf"><strong>Loop Engineering in the Harness: Autonomous, Self-Correcting Systems</strong></a> — Designing self-correcting Observe -&gt; Choose -&gt; Act -&gt; Verify loops with deterministic mechanical gates and stateless semantic critics.</li><li><a href="https://medium.com/google-cloud/from-vibing-chaos-to-reliable-agents-the-what-why-and-how-of-harness-engineering-6e7e0443a4fc"><strong>From Vibing Chaos to Reliable Agents: The What, Why, and How of Harness Engineering</strong></a> — How to wrap AI coding agents in architectural guardrails, linters, type-checkers, and automated verification harnesses.</li><li><a href="https://medium.com/google-cloud/skills-sprawl-when-too-much-of-a-good-thing-bloats-your-ai-agent-167f907e9dc2"><strong>Skills Sprawl: When Too Much of a Good Thing Confuses Your AI Agent</strong></a> — Why overloading your agent’s context window degrades reasoning, and how to partition and prune skills for peak accuracy.</li><li><a href="https://medium.com/google-cloud/documentation-as-context-a-skill-to-automate-your-blueprints-for-the-agentic-era-2bec0cf041a3"><strong>Documentation as Context: A Skill to Automate Your Blueprints for the Agentic Era</strong></a> — Keeping README, DESIGN, and architecture docs synchronised automatically so your coding agents never lose context.</li><li><a href="https://medium.com/google-cloud/give-your-ai-agents-deep-understanding-with-llms-txt-4f948590332b"><strong>Give Your AI Agents Deep Understanding with LLMs.txt (Part 1: Concept)</strong></a> — How curated llms.txt context files transform how agents understand complex repositories and documentation sets.</li><li><a href="https://medium.com/google-cloud/give-your-ai-agents-deep-understanding-creating-the-llms-txt-with-a-multi-agent-adk-solution-e5ae24bbd08b"><strong>Give Your AI Agents Deep Understanding — Creating a Multi-Agent ADK Solution: Design Phase (Part 2)</strong></a> — Designing a multi-agent system with Google ADK to synthesise deep repository context automatically.</li><li><a href="https://medium.com/google-cloud/give-your-ai-agents-deep-understanding-coding-the-multi-agent-adk-solution-041e844cdebb"><strong>Give Your AI Agents Deep Understanding — Coding the Multi-Agent ADK Solution (Part 3)</strong></a> — Full Python implementation of the multi-agent llms.txt generator using Google ADK.</li></ul><h3>2. Google Antigravity, MCP Servers &amp; Reusable Agent Skills</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*xWsJxXYvbf96gQGf.png" /></figure><p>Hands-on guides for <strong>Google Antigravity</strong>, building custom <strong>Model Context Protocol (MCP)</strong> integrations, and authoring portable agent skills:</p><ul><li><a href="https://medium.com/google-cloud/supercharging-the-gemini-app-with-custom-mcp-connectors-skills-and-live-github-access-d82c6cca5d85"><strong>Supercharging the Gemini App with Custom MCP Connectors, Skills, and Live GitHub Access</strong></a> — Extending the Gemini web and desktop experience with live external tools and GitHub integration.</li><li><a href="https://medium.com/google-cloud/meet-the-antigravity-extension-gemini-code-assist-on-steroids-33367e5a037c"><strong>Meet the Antigravity Extension: Gemini Code Assist on Steroids, in Whatever IDE You Love</strong></a> — Bringing agentic workflows and deep codebase awareness into your favourite editor.</li><li><a href="https://medium.com/google-cloud/dialling-our-agents-to-11-agent-skills-you-need-to-be-using-ccffa51e91df"><strong>Dialling Our Agents to 11: Agent Skills You Need to Be Using!</strong></a> — A curated tour of the highest-leverage agent skills for test-driven development, code review, and cloud automation.</li><li><a href="https://medium.com/google-cloud/dialling-our-agents-to-11-my-favourite-mcp-servers-9549c1442a5e"><strong>Dialling Our Agents to 11: My Favourite MCP Servers</strong></a> — The MCP servers that stay permanently enabled in my daily engineering workflow.</li><li><a href="https://medium.com/google-cloud/resolving-wsl-friction-with-google-antigravity-the-agy-2-0-and-agy-ide-edition-41cee17773c8"><strong>Resolving WSL Friction with Google Antigravity: the Agy 2.0 and Agy IDE Edition</strong></a> — Taming Windows Subsystem for Linux (WSL2) networking, browser attachment, and shell startup quirks alongside <a href="https://medium.com/google-cloud/working-with-google-antigravity-in-wsl-944c96c949f3"><strong>Working with Google Antigravity in WSL</strong></a>.</li><li><a href="https://medium.com/google-cloud/configuring-mcp-servers-and-skills-for-antigravity-cli-and-ide-a938c7eebb78"><strong>Configuring MCP Servers and Skills for Antigravity CLI and IDE</strong></a> — How to structure global and workspace configurations cleanly across tools.</li><li><a href="https://medium.com/google-cloud/confused-about-where-to-put-your-agent-skills-ea778f3c64f3"><strong>Confused About Where to Put Your Agent Skills?</strong></a> — Demystifying skill discovery paths, global vs workspace roots, and deduplication.</li><li><a href="https://medium.com/google-cloud/integrate-notebooklm-with-gemini-cli-google-antigravity-or-other-agents-with-mcp-cd83b575dc39"><strong>Integrate NotebookLM with Google Antigravity or Other Agents with MCP</strong></a> — Pairing <a href="https://medium.com/google-cloud/notebooklm-is-googles-insanely-cool-personal-ai-research-assistant-a2a36186f683"><strong>NotebookLM — Google’s Insanely Cool Personal AI Research Assistant</strong></a> directly with your coding agents.</li><li><a href="https://medium.com/google-cloud/helping-your-agent-understand-google-trying-the-developer-knowledge-api-and-mcp-server-425fc557b455"><strong>Help Your AI Understand Google: Trying the Developer Knowledge API and MCP Server</strong></a> — Eliminating hallucinated SDK methods and CLI flags by grounding agents in authoritative Google documentation.</li><li><a href="https://medium.com/google-cloud/power-up-your-gemini-cli-with-your-own-fastmcp-server-extension-75f1e78d0047"><strong>Power-Up Google Antigravity with Your Own FastMCP Server Extension</strong></a> — Building and packaging your own custom Python MCP server in minutes with FastMCP.</li><li><a href="https://medium.com/google-cloud/creating-an-automated-ui-test-of-your-web-app-in-seconds-with-gemini-cli-and-browsermcp-09cf4afb8940"><strong>Creating an Automated UI Test of Your Web App in Seconds, with Antigravity and BrowserMCP</strong></a> — Driving real browser sessions from your agent to verify frontend flows automatically.</li><li><a href="https://medium.com/google-cloud/trying-out-the-new-conductor-extension-in-gemini-cli-0801f892e2db"><strong>Trying Out the Conductor Extension — Adding Auth to Our Full Stack</strong></a> — Structured, spec-driven feature implementation on a full-stack application, alongside <a href="https://medium.com/google-cloud/using-gemini-cli-to-create-a-gemini-cli-config-repo-519399e25d9a"><strong>Creating a Version-Controlled Config Repo</strong></a> and <a href="https://medium.com/google-cloud/give-gemini-cli-the-ability-to-generate-images-and-video-work-with-github-repos-and-use-other-482172571f99"><strong>Generating Images, Video, and GitHub Workflows from the Terminal</strong></a>.</li></ul><h3>3. Multi-Agent Systems with Google ADK, Gemini &amp; the Rickbot Series</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*RWYdYkW7pb1KPiL_.png" /></figure><p>Building stateful, multi-agent applications on <strong>Google Cloud</strong> and <strong>Gemini Enterprise Agent Platform</strong> using the open-source <strong>Agent Development Kit (ADK)</strong>:</p><h3>Agentic Platforms &amp; FinSavant Series</h3><ul><li><a href="https://medium.com/google-cloud/finsavant-part-1-building-an-agentic-finops-platform-with-google-adk-a2ui-and-gemini-enterprise-248f59cea3a0"><strong>FinSavant Part 1: Building an Agentic FinOps Platform with Google ADK, A2UI and Gemini Enterprise Agent Platform</strong></a> — Architecture and design of a multi-agent cloud cost intelligence platform.</li><li><a href="https://medium.com/google-cloud/finsavant-part-2-building-an-agentic-finops-platform-development-environment-setup-google-dd12b8b84ba0"><strong>FinSavant Part 2: Development Environment Setup &amp; Google Cloud Integration</strong></a> — Wiring up the workspace, credentials, and foundational agent hierarchy.</li><li><a href="https://medium.com/google-cloud/building-a-multi-agent-finops-solution-with-google-adk-levering-tools-mcp-and-google-managed-assis-712bf7caa916"><strong>Building a Multi-Agent FinOps Solution with Google ADK, Leveraging Tools, MCP and Google-Managed Assistants</strong></a> — Combining custom ADK tools with managed cloud assistants to analyse and optimise spend.</li><li><a href="https://medium.com/google-cloud/agentic-firestore-smarter-agents-with-adk-and-google-remote-firestore-mcp-99c6c0240b5a"><strong>Agentic Firestore: Smarter Agents with ADK and Google Remote Firestore MCP</strong></a> — Connecting ADK agents directly to Google’s managed remote Firestore MCP endpoint with vector search (the exact architecture powering <a href="https://darrenlester.net">darrenlester.net</a>).</li><li><a href="https://medium.com/google-cloud/using-the-loop-pattern-to-make-my-multi-agent-solution-more-robust-86f8e9159a2a"><strong>Using the Loop Pattern to Make My Multi-Agent Solution More Robust (with Google ADK)</strong></a> — Implementing iterative validation and retry loops inside multi-agent orchestrations.</li></ul><h3>The Complete Rickbot Series (Gen AI SDK to Production ADK)</h3><ul><li><a href="https://medium.com/google-cloud/creating-a-rick-morty-chatbot-with-google-cloud-and-the-gen-ai-sdk-e8108e83dbee"><strong>Creating a Rick &amp; Morty Chatbot with Google Cloud and the Gen AI SDK</strong></a> — Kicking off the Rickbot journey with system instructions, persona tuning, and multimodal inputs.</li><li><a href="https://medium.com/google-cloud/adding-authentication-and-authorisation-to-our-rickbot-streamlit-chatbot-with-oauth-and-the-google-b892cda3f1d9"><strong>Adding Authentication and Authorisation to Our Rickbot Chatbot with OAuth</strong></a> — Securing the application with Google OAuth2 sign-in and session management.</li><li><a href="https://medium.com/google-cloud/building-the-rickbot-multi-personality-agentic-application-using-gemini-cli-google-a48aed4bef24"><strong>Building the Rickbot Multi-Personality Agentic Application (Part 1)</strong></a> — Evolving Rickbot into a multi-personality agentic system.</li><li><a href="https://medium.com/google-cloud/updating-the-rickbot-multi-personality-agentic-application-part-2-integrate-agent-development-ad39203e66ad"><strong>Updating the Rickbot Multi-Personality Application Part 2 — Integrating Google ADK</strong></a> — Migrating the core orchestration engine onto the Agent Development Kit (ADK).</li><li><a href="https://medium.com/google-cloud/updating-the-rickbot-multi-personality-agentic-application-part-3-guided-implementation-of-the-9675d3f92c11"><strong>Guided Implementation of Agent Development Kit (ADK) with Rickbot (Part 3)</strong></a> — Step-by-step walkthrough of ADK sessions, state, tools, and callbacks.</li><li><a href="https://medium.com/google-cloud/productionising-the-rickbot-adk-application-and-more-gemini-cli-tips-577cf6b37366"><strong>Productionising the Rickbot ADK Application</strong></a> — Hardening, containerising, and deploying the ADK agent to Google Cloud Run.</li><li><a href="https://medium.com/google-cloud/get-schwifty-with-the-fastapi-adding-a-rest-api-to-our-agentic-application-with-google-adk-6b87a4ea7567"><strong>Get Schwifty with FastAPI: Adding a REST API to Our Agentic Application (with Google ADK)</strong></a> — Exposing ADK agents cleanly over FastAPI endpoints for decoupled frontends.</li><li><a href="https://medium.com/google-cloud/introducing-google-adk-artifacts-for-multi-modal-file-handling-a-rickbot-blog-08ca6adf34c2"><strong>Introducing Google ADK Artifacts for Multi-Modal File Handling (a Rickbot Blog)</strong></a> — Handling images, PDFs, and binary blobs cleanly across agent turns using ADK Artifacts.</li><li><a href="https://medium.com/google-cloud/using-gemini-file-search-tool-for-rag-a-rickbot-blog-b6c4f117e5d3"><strong>Using Gemini File Search Tool for RAG (a Rickbot Blog)</strong></a> — Grounding ADK agents in custom document knowledge bases using Gemini’s native File Search RAG tool.</li></ul><h3>4. Enterprise Google Cloud Adoption, Migration &amp; Landing Zone Series</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*iaUImLT6aHir0NQL.png" /></figure><p>My comprehensive multi-part enterprise blueprint taking organisations from cloud strategy and business case through to Terraform Landing Zones, Migration Factories, and SRE operations:</p><h3>Phase 1: Cloud Strategy, Business Case &amp; Discovery</h3><ul><li><a href="https://medium.com/google-cloud/google-cloud-adoption-for-the-enterprise-from-strategy-to-operation-part-0-overview-9091f5a1ddfc"><strong>Google Cloud Adoption and Migration for the Enterprise: From Strategy to Operation — Series Overview (Part 0)</strong></a> — The end-to-end roadmap for enterprise cloud transformation.</li><li><a href="https://medium.com/google-cloud/google-cloud-adoption-for-the-enterprise-part-1a-strategy-eb2b863ff5f2"><strong>Establishing Your Cloud Strategy (Part 1A)</strong></a> — Defining business drivers, transformation principles, and target operating models.</li><li><a href="https://medium.com/@dazbo/google-cloud-adoption-for-the-enterprise-part-1b-reducing-software-license-costs-with-open-source-98453162ce8a"><strong>Reducing Software Licence Costs with Open Source (Part 1B)</strong></a> — Escaping punitive commercial licensing traps by modernising onto open-source and cloud-native engines.</li><li><a href="https://medium.com/@dazbo/data-centre-exit-mainframe-exit-and-byod-google-cloud-adoption-1c-bc441e16c70f"><strong>Cloud Adoption: Data Centre Exit, Mainframe Exit, and BYOD (Part 1C)</strong></a> — Navigating lease expiries, hardware refresh cycles, and mainframe offloading.</li><li><a href="https://medium.com/@dazbo/lift-and-shift-gce-and-gcve-google-cloud-adoption-1d-aec728aa936c"><strong>Lift-and-Shift, GCE and GCVE (Part 1D)</strong></a> — When to rehost on Compute Engine or Google Cloud VMware Engine (GCVE) versus refactoring.</li><li><a href="https://medium.com/google-cloud/strategy-checklist-google-cloud-adoption-series-1e-9399c1ca3260"><strong>Cloud Strategy Checklist (Part 1E)</strong></a> — The executive sanity check before committing to execution.</li><li><a href="https://medium.com/@dazbo/creating-your-cloud-adoption-business-case-google-cloud-adoption-part-2-c986cca2505d"><strong>Creating Your Cloud Adoption Business Case (Part 2)</strong></a> — Building a defensible TCO and value model that wins CFO approval.</li><li><a href="https://medium.com/google-cloud/cloud-adoption-and-cloud-engineering-principles-google-cloud-adoption-part-3-660bdb78cebb"><strong>Cloud Adoption and Cloud Engineering — Architecture Principles (Part 3)</strong></a> — Establishing non-negotiable architectural tenets for cloud-native engineering.</li><li><a href="https://medium.com/@dazbo/discovery-and-mapping-your-as-is-to-cloud-and-using-cloud-migraiton-center-google-cloud-adoption-3253a11f18b8"><strong>Discovery and Mapping Your As-Is to Cloud, and Using Cloud Migration Center</strong></a> — Running automated estate discovery and dependency mapping with Google Cloud Migration Center.</li><li><a href="https://medium.com/google-cloud/mapping-your-as-is-to-google-cloud-to-leverage-cloud-value-6609b2368678"><strong>Mapping Your As-Is to Google Cloud to Leverage Cloud Value</strong></a> &amp; <a href="https://medium.com/google-cloud/mapping-as-is-to-leverage-cloud-value-os-middleware-and-dbs-3002a80cd709"><strong>Mapping to Maximise Cloud Value: OS, Application Servers and Databases</strong></a> — Choosing the right modernisation disposition (6 Rs) for every layer of your stack.</li></ul><h3>Phase 2: Landing Zone Architecture, Terraform FAST &amp; Governance</h3><ul><li><a href="https://medium.com/google-cloud/landing-zones-on-google-cloud-b42b08e1abaa"><strong>Landing Zones on Google Cloud: What It Is, Why You Need One, and How to Create One</strong></a> — The foundational introduction to Google Cloud Landing Zones.</li><li><a href="https://medium.com/google-cloud/google-cloud-landing-zone-with-terraform-and-cloud-foundation-fabric-fast-part-1-2c7634ac31fd"><strong>Google Cloud Landing Zone with Terraform and Cloud Foundation Fabric FAST — Part 1</strong></a> &amp; <a href="https://medium.com/google-cloud/google-cloud-landing-zone-with-terraform-and-cloud-foundation-fabric-fast-part-2-6bd067f5bed7"><strong>FAST — Part 2</strong></a> — Hands-on deployment of Google’s Enterprise Foundations blueprint using Cloud Foundation Fabric FAST.</li><li><a href="https://medium.com/google-cloud/design-your-landing-zone-design-considerations-part-1-google-cloud-adoption-series-74978d968fed"><strong>Design Your Landing Zone — Part 1: Core Foundations</strong></a> — Resource hierarchy, Cloud Identity, IAM, Shared VPC networking, and hybrid connectivity.</li><li><a href="https://medium.com/google-cloud/design-your-landing-zone-design-considerations-part-2-kubernetes-and-gke-google-cloud-5a500384cb03"><strong>Design Your Landing Zone — Part 2: Kubernetes and GKE</strong></a> — Multi-tenant GKE cluster topology, fleet management, and IP address planning.</li><li><a href="https://medium.com/google-cloud/design-your-landing-zone-design-considerations-part-3-monitoring-logging-billing-and-7b40189a3c81"><strong>Design Your Landing Zone — Part 3: Monitoring, Logging, Billing and Security</strong></a> — Centralised observability, FinOps billing exports, Security Command Center, and KMS.</li><li><a href="https://medium.com/google-cloud/design-your-landing-zone-design-considerations-part-4-iac-gitops-and-ci-cd-google-cloud-ae3f533c6dbd"><strong>Design Your Landing Zone — Part 4: IaC, GitOps and CI/CD</strong></a> — Structuring Terraform repositories, state management, and automated policy validation.</li><li><a href="https://medium.com/google-cloud/landing-zone-technical-onboarding-the-how-to-google-cloud-adoption-series-9b7ba8710e83"><strong>Landing Zone Technical Onboarding — The “How-To”</strong></a> &amp; <a href="https://medium.com/google-cloud/landing-zone-deployment-google-cloud-adoption-series-bdc0b36106d0"><strong>Landing Zone Deployment</strong></a> — Running technical onboarding workshops and executing the production rollout.</li><li><a href="https://medium.com/google-cloud/i-deleted-a-bunch-of-org-policies-heres-what-google-cloud-told-me-390b98a9a5c4"><strong>I Deleted a Bunch of Org Policies! Here’s What Google Cloud Told Me…</strong></a> — What actually happens when Organisation Policy constraints are removed or overridden across folders and projects.</li><li><a href="https://medium.com/google-cloud/cloud-adoption-enabling-your-consumers-and-tenants-and-the-project-factory-d868f4b1443a"><strong>Cloud Adoption: Enabling Your Consumers and Tenants, and the Project Factory</strong></a> — Automating self-service project vending with guardrails baked in.</li></ul><h3>Phase 3: Operating Model, SRE &amp; Migration Execution at Scale</h3><ul><li><a href="https://medium.com/@dazbo/google-cloud-adoption-organisational-change-capabilities-upskilling-and-cloud-centre-of-15bc49ae7ae6"><strong>Organisational Change, Capabilities, Upskilling, and Cloud Centre of Excellence (CCoE)</strong></a> — Building the human capabilities and CCoE structures that make cloud adoption stick.</li><li><a href="https://medium.com/google-cloud/google-cloud-adoption-site-reliability-engineering-sre-and-best-practices-for-sli-slo-sla-6670c864c96b"><strong>Site Reliability Engineering (SRE), and Best Practices for SLI / SLO / SLA</strong></a> — Implementing practical error budgets, SLIs, and SLOs in enterprise operations.</li><li><a href="https://medium.com/google-cloud/how-to-accelerate-cloud-migrations-at-scale-ecee08a09089"><strong>How to Accelerate Cloud Migrations at Scale</strong></a>, <a href="https://medium.com/google-cloud/accelerating-cloud-migration-with-the-migration-factory-and-migration-tools-e30d3d4884db"><strong>Accelerating Cloud Migration with the Migration Factory and Migration Tools</strong></a>, and <a href="https://medium.com/google-cloud/what-should-you-migrate-to-the-cloud-and-in-what-order-5149fec9a5b9"><strong>Sequencing Cloud Migration to Reduce Cost: What to Migrate and When</strong></a> — Running wave-based migration factories without blowing the transformation budget.</li></ul><h3>5. Cloud FinOps, Security, Serverless &amp; GitOps Automation</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*zLnhtpyEN8yAe0HS.png" /></figure><p>Practical guides for keeping cloud spend under control, securing Cloud Run services, and automating CI/CD pipelines:</p><ul><li><a href="https://medium.com/google-cloud/finally-hard-caps-to-limit-your-google-cloud-spend-29b2e658ed73"><strong>Finally — Hard Caps to Limit Your Google Cloud Spend</strong></a> — Using Google Cloud’s budget enforcement capabilities to prevent runaway billing shocks.</li><li><a href="https://medium.com/google-cloud/how-to-avoid-a-massive-cloud-bill-41a76251caba"><strong>Automated GCP Killswitch: How to Avoid an Unexpected Google Cloud Bill</strong></a> — Event-driven Cloud Run / Cloud Functions circuit breakers that automatically cap runaway projects.</li><li><a href="https://medium.com/google-cloud/using-google-identity-aware-proxy-iap-with-cloud-run-without-a-load-balancer-27db89b9ed49"><strong>Using Google Identity-Aware Proxy (IAP) with Cloud Run — Without a Load Balancer!</strong></a> — Direct, zero-trust IAP protection on Cloud Run services without paying for a Global External Application Load Balancer.</li><li><a href="https://medium.com/google-cloud/automated-github-code-reviews-using-google-gemini-update-enterprise-ready-multimodal-and-e3459f621aea"><strong>Automated GitHub Code Reviews Using Google Gemini Update: Enterprise-Ready, Multimodal, and Smart</strong></a> &amp; <a href="https://medium.com/google-cloud/automated-github-code-reviewsusing-google-gemini-7b4d027b3092"><strong>Original Guide</strong></a> — My open-source gemini-review-action GitHub Action for automated PR reviews and issue triage.</li><li><a href="https://medium.com/google-cloud/create-a-ci-cd-pipeline-using-github-actions-and-google-cloud-9be20ff50e97"><strong>Create a CI/CD Pipeline Using GitHub Actions and Google Cloud</strong></a> &amp; <a href="https://medium.com/google-cloud/integrating-our-application-ci-cd-pipelines-and-terraform-gitops-with-cloud-build-35e8d38b8468"><strong>Integrating Application CI/CD Pipelines and Terraform GitOps with Cloud Build</strong></a> — End-to-end infrastructure and application delivery pipelines.</li><li><a href="https://medium.com/google-cloud/a-sample-application-using-terraform-on-google-cloud-2405c598a60a"><strong>A Sample Application Using Terraform on Google Cloud: Design and Implementation</strong></a> &amp; <a href="https://medium.com/google-cloud/creating-a-static-website-using-google-cloud-storage-814c3b90ec8a"><strong>Creating a Static Website Using Google Cloud Storage</strong></a> — Clean reference patterns for serverless and static hosting on GCP.</li></ul><h3>6. AI Video Intelligence, Media Processing &amp; Pre-Trained AI</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*-uANF6Yc-HszKa3g.png" /></figure><p>Building serverless Python applications that download, transcribe, translate, and analyse video and image content on Google Cloud Run:</p><ul><li><a href="https://python.plainenglish.io/downloading-youtube-videos-extracting-audio-and-generating-transcripts-with-python-and-jupyter-c3068f82bbe0"><strong>Downloading YouTube Videos, Extracting Audio, and Generating Transcripts with Python and Jupyter</strong></a> — Part 1: Local audio extraction and speech-to-text transcription pipelines.</li><li><a href="https://python.plainenglish.io/youtube-video-downloader-with-generative-ai-and-python-run-anywhere-transcribe-and-translate-dec2e593dd58"><strong>YouTube Video Downloader with Generative AI and Python: Run Anywhere, Transcribe and Translate</strong></a> — Part 2: Adding Gemini-powered summarisation and multi-language translation.</li><li><a href="https://medium.com/google-cloud/running-ai-youtube-and-video-processing-as-a-python-streamlit-web-application-and-hosting-on-748aae8e54b4"><strong>Building and Running AI YouTube and Video Processing on Cloud Run</strong></a> — Part 3: Packaging the pipeline as a Streamlit web application hosted on Google Cloud Run.</li><li><a href="https://medium.com/google-cloud/deploying-our-video-intelligence-cloud-run-appication-with-terraform-and-cloud-build-a1f4dac7ba6a"><strong>Deploying Our Video Intelligence Cloud Run Application with Terraform and Cloud Build</strong></a> &amp; <a href="https://medium.com/google-cloud/optimising-the-cost-of-our-video-intelligence-service-on-google-cloud-run-aided-with-billing-a9985c6208c6"><strong>Optimising the Cost of Our Video Intelligence Service Aided with Billing Reports</strong></a> — Automating deployment and tuning CPU/memory allocation using real GCP billing data.</li><li><a href="https://medium.com/google-cloud/building-a-serverless-image-text-extractor-and-translator-using-google-cloud-pre-trained-ai-adfdccdb18d9"><strong>Building a Serverless Image Text Extractor and Translator Using Google Cloud Pre-Trained AI</strong></a> — Event-driven OCR and translation using Cloud Vision, Translation API, Eventarc, and Cloud Run.</li></ul><h3>7. Architecture Leadership, Google Cloud Certifications &amp; Challenge Labs</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*dPggCLFi-PF42JJZ.png" /></figure><p>Practical advice on enterprise architecture, communication, passing Google Cloud Professional certifications, and solving hands-on Skill Badge Challenge Labs:</p><h3>Architecture Craft &amp; Event Recaps</h3><ul><li><a href="https://medium.com/@dazbo/architecture-documentation-where-to-draw-the-line-df73fb5ca85c"><strong>Architecture Documentation: Where to Draw the Line?</strong></a> — How much detail is “just enough” in HLD/LLD documents, and why living boundaries beat 200-page shelfware.</li><li><a href="https://medium.com/@dazbo/the-unwritten-rules-of-it-architecture-from-imposter-to-impact-ac428d1d692d"><strong>The Unwritten Rules of IT Architecture: From Imposter to Impact</strong></a> — Lessons from decades in enterprise technology on stakeholder alignment, trade-offs, and technical leadership.</li><li><a href="https://medium.com/@dazbo/compelling-pitches-and-arguments-for-any-scenario-top-down-thinking-108c9edda6ec"><strong>Compelling Pitches and Arguments for Any Scenario: Top-Down Thinking</strong></a> — Using the Minto Pyramid Principle and structured storytelling to win executive buy-in.</li><li><a href="https://medium.com/@dazbo/my-incredible-google-cloud-next-26-experience-as-a-gde-1bd291cc3696"><strong>My Incredible Google Cloud Next ’26 Experience as a GDE</strong></a> — Highlights and key announcements from Las Vegas, alongside my earlier recaps from <a href="https://medium.com/@dazbo/cloud-ai-infrastructure-show-2025-165ce2019ab8"><strong>Cloud &amp; AI Infrastructure Show 2025</strong></a>, <a href="https://medium.com/google-cloud/dazbos-google-cloud-next-24-recap-keynote-6f5518238c9d"><strong>Google Cloud Next ’24 Keynote</strong></a>, <a href="https://medium.com/@dazbo/cloud-expo-recap-2024-9ecb4ed26cd7"><strong>Cloud Expo 2024</strong></a>, <a href="https://medium.com/google-cloud/google-next-2023-experience-and-favourite-sessions-fb00add5f59e"><strong>Google Cloud Next </strong></a>’23, and <a href="https://medium.com/@dazbo/google-cloud-next-22-summaries-1f703d511b65"><strong>Google Cloud Next ‘22</strong></a>.</li></ul><h3>Google Cloud Certification Exam Guides &amp; Challenge Lab Walkthroughs</h3><ul><li><strong>Professional Certification Guides:</strong> <a href="https://medium.com/google-cloud/how-to-pass-the-google-professional-cloud-devops-exam-fc8754af0203"><strong>How to Pass the Google Professional Cloud DevOps Engineer Exam</strong></a>, <a href="https://medium.com/google-cloud/tips-for-passing-the-google-professional-developer-pcd-exam-c2b6d2253d17"><strong>Tips for Passing the Google Professional Cloud Developer (PCD) Exam</strong></a>, and <a href="https://medium.com/google-cloud/tips-for-passing-or-repassing-the-google-professional-data-engineer-pcde-exam-53860d358e8e"><strong>Tips for Passing (or Repassing!) the Google Professional Data Engineer (PCDE) Exam</strong></a>.</li><li><a href="https://medium.com/@dazbo/google-cloud-challenge-lab-walkthroughs-77a1acf52848"><strong>Google Cloud Challenge Lab Walkthroughs — Master Hub</strong></a> — Index of step-by-step solutions, including <a href="https://medium.com/google-cloud/build-infrastructure-on-google-cloud-with-terraform-google-challenge-lab-walkthrough-30a592373d3e"><strong>Build Infrastructure with Terraform</strong></a>, <a href="https://medium.com/google-cloud/develop-genai-apps-with-gemini-and-streamlit-google-cloud-challenge-lab-walkthrough-a3bedaea7cee"><strong>Develop GenAI Apps with Gemini and Streamlit</strong></a>, <a href="https://medium.com/google-cloud/google-cloud-challenge-lab-implement-devops-4e3463a4fe59"><strong>Implement DevOps in Google Cloud</strong></a>, <a href="https://medium.com/google-cloud/set-up-and-configure-a-google-cloud-environment-google-challenge-lab-walkthrough-cb2a18218ebd"><strong>Set Up and Configure a Google Cloud Environment</strong></a>, <a href="https://medium.com/google-cloud/engineer-data-in-google-cloud-google-cloud-challenge-lab-walkthrough-06333d14685b"><strong>Engineer Data in Google Cloud</strong></a>, <a href="https://medium.com/google-cloud/develop-serverless-applications-on-cloud-run-google-cloud-challenge-lab-walkthrough-a62da1c33989"><strong>Develop Serverless Applications on Cloud Run</strong></a>, <a href="https://medium.com/@dazbo/deploy-kubernetes-applications-on-google-cloud-google-cloud-challenge-lab-walkthrough-ff4ecf170108"><strong>Deploy Kubernetes Applications on Google Cloud</strong></a>, <a href="https://medium.com/google-cloud/optimise-costs-for-google-kubernetes-engine-google-cloud-challenge-lab-walkthrough-739abfb58ea0"><strong>Optimise Costs for GKE</strong></a>, <a href="https://medium.com/google-cloud/configuring-vpn-connectivity-between-google-cloud-and-aws-google-cloud-challenge-lab-walkthrough-aa2266422078"><strong>Configure VPN Connectivity Between Google Cloud and AWS</strong></a>, and <a href="https://medium.com/@dazbo/develop-serverless-apps-with-firebase-google-cloud-challenge-lab-walkthrough-1297e33d8708"><strong>Develop Serverless Apps with Firebase</strong></a>.</li></ul><h3>8. Python, Advent of Code, Hardware &amp; Developer Productivity</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*N-tmGUOzuFKZg6p5.png" /></figure><p>Tutorials on Python packaging, symbolic mathematics with SymPy, Advent of Code solution walkthroughs, and practical hardware/OS hacks:</p><ul><li><strong>Advent of Code &amp; Python Mastery:</strong> <a href="https://python.plainenglish.io/advent-of-code-the-best-way-to-learn-to-code-d86c6478d484"><strong>Advent of Code: The Best Way to Learn to Code!</strong></a> and <a href="https://medium.com/@dazbo/advent-of-code-2022-4a4b82d4f87"><strong>Advent of Code: Python Solutions and Walkthroughs</strong></a> (companion to my full walkthrough site at <a href="https://aoc.just2good.co.uk/"><strong>aoc.just2good.co.uk</strong></a>).</li><li><strong>Symbolic Maths &amp; Jupyter in Python:</strong> <a href="https://python.plainenglish.io/solving-mathematical-problems-in-python-with-sympy-5f138c0deaef"><strong>Solving Mathematical Problems in Python with SymPy</strong></a>, <a href="https://python.plainenglish.io/python-sympy-to-solve-linear-equations-238cbc733771"><strong>Python SymPy to Solve Linear Equations</strong></a>, <a href="https://python.plainenglish.io/finding-the-coefficients-of-a-quadratic-with-sympy-777cd9479ea1"><strong>Finding the Coefficients of a Quadratic with SymPy</strong></a>, <a href="https://python.plainenglish.io/five-ways-to-run-jupyter-labs-and-notebooks-23209f71e5c0"><strong>Six Ways to Run Jupyter Labs and Notebooks</strong></a>, and <a href="https://python.plainenglish.io/noobs-guide-to-creating-a-pip-installable-distributable-python-package-ed52e03cd4e8"><strong>Noob’s Guide to Creating a Pip-Installable Distributable Python Package</strong></a>.</li><li><strong>Developer Hardware, Windows &amp; Web Tooling:</strong> <a href="https://medium.com/@dazbo/should-you-switch-to-a-windows-arm-laptop-my-experience-of-the-surface-laptop-7-f43725133f24"><strong>Should You Switch to a Windows ARM Laptop? (Surface Laptop 7)</strong></a>, <a href="https://medium.com/@dazbo/upgrading-your-microsoft-surface-laptop-hard-disk-easier-than-you-think-543a11b1c46f"><strong>Upgrading Your Microsoft Surface Laptop Hard Disk — Easier Than You Think!</strong></a>, <a href="https://medium.com/@dazbo/dual-computer-dual-monitor-with-kvm-1bad7ac10b7a"><strong>How to Configure Two Computers and Two Monitors with a Cheap KVM</strong></a>, <a href="https://medium.com/@dazbo/wingetui-awesome-package-manager-ui-for-windows-f902d2f4a5b8"><strong>WingetUI — Awesome Package Manager UI for Windows</strong></a>, <a href="https://medium.com/@dazbo/create-a-long-full-page-screenshot-from-chrome-0ab2673534df"><strong>Create a Long / Full-Page Screenshot from Chrome</strong></a>, <a href="https://medium.com/@dazbo/vpn-blocked-on-public-wifi-5e9e8ee468a5"><strong>Is Your VPN Blocked on Public Wi-Fi? Here’s an Easy Fix</strong></a>, <a href="https://medium.com/@dazbo/jekyll-container-image-a7785f9c5ed6"><strong>Creating Websites with Jekyll and a Pre-Canned Container Image</strong></a>, <a href="https://medium.com/@dazbo/cross-posting-from-wordpress-to-medium-and-to-other-social-media-platforms-4a7904f737ab"><strong>Cross-Posting from WordPress to Medium</strong></a>, <a href="https://medium.com/@dazbo/25000-hack-attempts-time-to-enable-2fa-for-my-wordpress-site-53a18ae03262"><strong>25,000 Hack Attempts? Time to Enable 2FA for WordPress</strong></a>, <a href="https://medium.com/@dazbo/asking-chatgpt-about-its-processors-and-google-cloud-e785c0581755"><strong>Asking ChatGPT About Its Processors and Google Cloud</strong></a>, and (for the petrolheads!) <a href="https://medium.com/@dazbo/going-from-a-2017-mustang-gt-5-0-to-a-2020-bed13c117fb4"><strong>Going from a 2017 Mustang GT 5.0 to a 2020</strong></a>.</li></ul><h3>Readers’ Top 10 Most-Read Articles (520K+ Lifetime Views)</h3><p>Here are the ten most-read guides across my entire Medium catalogue:</p><ul><li><a href="https://medium.com/@dazbo/create-a-long-full-page-screenshot-from-chrome-0ab2673534df"><strong>Create a Long / Full-Page Screenshot from Chrome (for Full-Screen Copy-and-Paste)</strong></a> — Over 103,000 views: the quickest built-in Chrome DevTools trick for capturing full-length scrolling web pages without sketchy browser extensions.</li><li><a href="https://medium.com/google-cloud/working-with-google-antigravity-in-wsl-944c96c949f3"><strong>Working with Google Antigravity in WSL</strong></a> — Over 43,000 views: the original deep dive into configuring Google Antigravity cleanly inside Windows Subsystem for Linux (WSL2).</li><li><a href="https://medium.com/@dazbo/should-you-switch-to-a-windows-arm-laptop-my-experience-of-the-surface-laptop-7-f43725133f24"><strong>Should You Switch to a Windows ARM Laptop? My Experience of the Surface Laptop 7</strong></a> — Over 40,000 views: an honest, developer-focused review of running Python, WSL, Docker, and daily engineering workflows on Snapdragon X Elite ARM hardware.</li><li><a href="https://medium.com/@dazbo/dual-computer-dual-monitor-with-kvm-1bad7ac10b7a"><strong>How to Configure Two Computers and Two Monitors with a Cheap KVM</strong></a> — Over 29,000 views: how to wire up a dual-computer, dual-monitor desk setup without spending hundreds on specialist KVM switches.</li><li><a href="https://medium.com/@dazbo/vpn-blocked-on-public-wifi-5e9e8ee468a5"><strong>Is Your VPN Blocked on Public Wi-Fi? Here’s an Easy Fix</strong></a> — Over 27,000 views: a fast, practical workaround when hotel, train, or guest Wi-Fi networks block standard VPN protocols.</li><li><a href="https://medium.com/google-cloud/configuring-mcp-servers-and-skills-for-antigravity-cli-and-ide-a938c7eebb78"><strong>Configuring MCP Servers and Skills for Antigravity CLI and IDE</strong></a> — Over 15,000 views: the definitive guide to wiring up Model Context Protocol (MCP) servers and reusable agent skills.</li><li><a href="https://medium.com/google-cloud/integrate-notebooklm-with-gemini-cli-google-antigravity-or-other-agents-with-mcp-cd83b575dc39"><strong>Integrate NotebookLM with Google Antigravity or Other Agents with MCP</strong></a> — Over 11,000 views: connecting NotebookLM’s grounded research engine directly into your coding agent via MCP.</li><li><a href="https://python.plainenglish.io/downloading-youtube-videos-extracting-audio-and-generating-transcripts-with-python-and-jupyter-c3068f82bbe0"><strong>Downloading YouTube Videos, Extracting Audio, and Generating Transcripts with Python and Jupyter</strong></a> — Over 10,500 views: building an automated media extraction and AI transcription pipeline in Python.</li><li><a href="https://medium.com/google-cloud/create-a-ci-cd-pipeline-using-github-actions-and-google-cloud-9be20ff50e97"><strong>Create a CI/CD Pipeline Using GitHub Actions and Google Cloud</strong></a> — Over 9,300 views: end-to-end automated build, test, and deployment pipelines to Google Cloud using GitHub Actions.</li><li><a href="https://medium.com/google-cloud/design-your-landing-zone-design-considerations-part-1-google-cloud-adoption-series-74978d968fed"><strong>Design Your Landing Zone — Design Considerations Part 1</strong></a> — Over 8,800 views: resource hierarchy, identity, networking, and security foundations for enterprise Google Cloud environments.</li></ul><h3>Where to Find Everything Else</h3><p>If you want to explore my open-source projects, watch video walkthroughs, or try out live applications:</p><ul><li><strong>Portfolio &amp; AI Assistant:</strong> <a href="https://darrenlester.net">https://darrenlester.net</a></li><li><strong>Full Searchable Blog Catalogue:</strong> <a href="https://darrenlester.net/blogs">https://darrenlester.net/blogs</a></li><li><strong>Dev.to Articles:</strong> <a href="https://dev.to/dazbo">https://dev.to/dazbo</a></li><li><strong>Open-Source Projects &amp; Repositories:</strong> <a href="https://darrenlester.net/projects">https://darrenlester.net/projects</a> and <a href="https://github.com/derailed-dash">GitHub (</a><a href="https://github.com/derailed-dash">@derailed-dash)</a></li><li><strong>Advent of Code Walkthroughs &amp; Python Learning:</strong> <a href="https://aoc.just2good.co.uk/">https://aoc.just2good.co.uk/</a></li><li><strong>LinkedIn:</strong> <a href="https://www.linkedin.com/in/darren-lester-architect/">Darren Lester — Google Chief Technologist, Google Developer Expert (GDE), Cloud Ambassador and Enterprise Cloud Architect</a></li><li><strong>Get Email Updates:</strong> <a href="https://medium.com/@dazbo/subscribe">Subscribe to </a><a href="https://medium.com/@dazbo/subscribe">@dazbo on Medium</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cd21558dde32" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Supercharging the Gemini App with Custom MCP Connectors, Skills, and Live GitHub Access]]></title>
            <link>https://medium.com/google-cloud/supercharging-the-gemini-app-with-custom-mcp-connectors-skills-and-live-github-access-d82c6cca5d85?source=rss-58a0bd8ba726------2</link>
            <guid isPermaLink="false">https://medium.com/p/d82c6cca5d85</guid>
            <category><![CDATA[mcp-server]]></category>
            <category><![CDATA[gemini]]></category>
            <category><![CDATA[google-ai-studio]]></category>
            <category><![CDATA[github]]></category>
            <category><![CDATA[antigravity]]></category>
            <dc:creator><![CDATA[Dazbo (Darren Lester)]]></dc:creator>
            <pubDate>Wed, 04 Nov 2026 02:38:15 GMT</pubDate>
            <atom:updated>2026-11-04T02:38:15.072Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*K8U4Sbm8ROnZ_XqraJWgwQ.png" /><figcaption>Dazbo picking a Google AI surface</figcaption></figure><p>Howdy friends! If you spend as much time playing with Google’s AI ecosystem as I do, you’re probably jumping between different Google tools frequently. Each tool plays a distinct role in your workflow.</p><p>For me, almost all of my serious work — whether creating applications, building sites, or writing blogs — happens inside <strong>Google Antigravity</strong>. Empowered with context, skills, and MCP servers, Agy is the agentic harness that knows my expectations, rules, and all of my GitHub repos — including my dazbo-content writing persona. It&#39;s basically my agentic ghost writer!</p><p>But I also use the <strong>Gemini Web App</strong> (gemini.google.com) and mobile app as my everyday Swiss Army knife. Whether I&#39;m proofreading a LinkedIn post, planning flights, knocking up a video, or asking Gemini to airbrush a rogue pizza box out of a photo, the Gemini App is always ready in a browser tab.</p><p>There was just one friction point: my GitHub context was not available to the Gemini App. My writing persona, reusable skills, and project docs live inside my public and private GitHub repositories. I wanted the Gemini App to have on-demand access to <em>all</em> these repos across <em>every</em> conversation, without manually copy-pasting files like a caveman.</p><blockquote>“Hang on, Dazbo — doesn’t the Gemini app already have a built-in GitHub connector?”</blockquote><blockquote>“It does indeed.”</blockquote><blockquote>“So you just toggled it on and went to the pub?”</blockquote><blockquote>“If only it were that simple.”</blockquote><p>As I quickly discovered, the out-of-the-box GitHub connector in the Gemini web app is not quite what you think it is. Fortunately, Google has added a couple of powerhouse customisation features to the Gemini app: <strong>Custom Model Context Protocol (MCP) Connectors</strong> and persistent <strong>Skills</strong>.</p><p>In this article, I’ll disambiguate where the Gemini App fits alongside Google AI Studio and Antigravity, tour the pre-made connectors, see why the native GitHub connector falls short, and walk step-by-step through wiring up GitHub’s managed remote MCP server.</p><h3>Wait — Which Tool Should I Use?</h3><p>Google drops a lot of products with overlapping capabilities, and their branding can be <em>so much fun</em> to keep up with. Here is how I use these tools in my own workflows:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*L2t6lQhY0lsHQLW6Y9OAiQ.png" /><figcaption>Google AI Tools Ecosystem</figcaption></figure><h4>1. <a href="https://antigravity.google/?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Google Antigravity</a> (Antigravity Extension for Your Favourite IDE, Antigravity IDE, &amp; Antigravity CLI)</h4><p>Antigravity is Google’s agentic development harness. It’s my daily driver for anything I do that is worthy of its own Git repo. This even includes writing blogs and whitepapers.</p><h4>2. Google AI Studio (<a href="https://aistudio.google.com?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">aistudio.google.com</a>)</h4><p>If you haven’t looked closely at <strong>Google AI Studio</strong> lately, you might still think of it purely as a prompt playground where you tweak temperature sliders and grab a GEMINI_API_KEY. It still does that; but here&#39;s what else you can do with it:</p><ul><li>From the <strong>Playground</strong>, explore the latest Google models — such as Gemini Argon, Gemini Omni, Nano Banana, TTS, Lyria, Live Translate, and Gemma — and toggle grounding with Google Search or Google Maps.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/805/1*ZMMu2-_vripJEeMyXnuAlw.png" /><figcaption>Model selection in the Playground</figcaption></figure><ul><li>Configure models in the UI and open ready-to-execute code snippets in one click. This screenshot shows how easy it is to obtain the code for low-latency real-time audio conversation using gemini-3.8-live, or open it directly in <a href="https://colab.research.google.com/?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Google Colab</a>.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nb9Wrfno7S3JKgFH6NhaAQ.png" /><figcaption>Code snippets from AI Studio</figcaption></figure><ul><li><a href="https://aistudio.google.com/apps?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco"><strong>Build</strong></a> new applications from a prompt.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nMORezIIQWm-EW94148Z5Q.png" /><figcaption>Build with AI Studio</figcaption></figure><ul><li>Explore the <a href="https://aistudio.google.com/apps?source=showcase&amp;showcaseTag=featured&amp;utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco"><strong>Gallery</strong></a> of pre-built agents and applications to use as-is, or <em>remix</em> into your own apps.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5vuxTCNVlhPh-ZXWxUOPow.png" /><figcaption>Gallery in AI Studio</figcaption></figure><ul><li>Use the <a href="https://aistudio.google.com/usage?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco"><strong>Dashboard</strong></a> to manage API keys, link Google Cloud projects, view model usage, or set hard monthly spend limits.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AZGHEf2GxVd4zebEXEoQDQ.png" /><figcaption>AI Studio Dashboard</figcaption></figure><ul><li><strong>One-Click <em>Cloud Run</em> Publishing, GitHub Sync &amp; Direct Export to Antigravity:</strong> <em>(Tip: If you open a showcase app from the </em><strong><em>Gallery</em></strong><em>, it initially opens in read-only mode — click </em><strong><em>Remix</em></strong><em> in the top-right corner first to create your own editable copy!)</em> Once you own the app, click <strong>Publish</strong> in the top-right header to deploy it straight to Google serverless <strong>Cloud Run</strong>. And what if that prototype needs serious engineering — maybe an automated test suite? Switch to the <strong>&lt;&gt; Code</strong> tab and click <strong>Export</strong> in the top-right corner to <strong>Push to GitHub</strong> or click <strong>Export to Antigravity</strong> to open it directly in <strong>Google Antigravity</strong>!</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HctClXxfAYhv5JUPv6tb_Q.png" /><figcaption>Export to</figcaption></figure><p>So clearly, AI Studio is an amazing way to rapidly build and deploy an application from scratch. But it’s also great for creating a prototype or seeding a project before working on it in Antigravity.</p><h4>3. The Gemini App (<a href="https://gemini.google.com?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">gemini.google.com</a>) &amp; Mobile App</h4><p>This is your web-based conversational AI assistant and research workspace. It has no access to your local file system or tools.</p><p>Alongside text chats, you can have real-time voice and live camera/screen-sharing video conversations with your assistant, plus:</p><ul><li><strong>Multimodal image, audio &amp; video creation, and in-place editing.</strong></li><li>Integration with Google <strong>Gemini Notebook</strong> (formerly NotebookLM).</li><li>Integration with the <strong>Deep Research Agent</strong>.</li><li>Integration with <strong>Canvas</strong>: A side-by-side workspace for co-writing documents and slide decks (with inline tone/length sliders and one-click export to Google Docs or Slides), transforming notes into infographics, quizzes, or Audio Overviews, and live-previewing interactive HTML/React apps and 3D visualisations (or exporting Python scripts to Google Colab).</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ANAvTtmD0KGq76gJHhhGcA.png" /><figcaption>Gemini App</figcaption></figure><ul><li><a href="https://gemini.google.com/scheduled?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco"><strong>Scheduled Actions</strong></a>: Built-in background automation that runs one-off or recurring prompts on a schedule — such as a daily morning tech briefing or weekly project summary.</li><li>Persistent <strong>Skills</strong>: Install skills into the Gemini App and invoke them explicitly with @ or /, or implicitly by letting Gemini activate matching skills automatically. Because these use the open SKILL.md standard, you can share the exact same skill folders with <strong>Google Antigravity</strong> (or any other coding agent).</li><li><a href="https://gemini.google.com/customize/connectors?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco"><strong>Connected Apps / Connectors</strong></a>: Links your assistant to first-party Google services (Google Workspace, Gemini Notebook, YouTube Music), over 40 partner integrations (Canva, Adobe, Dropbox, Shopify, monday.com), and custom Model Context Protocol (MCP) servers.</li></ul><p>In my own day-to-day workflow, it’s my favourite tool for anything that doesn’t warrant a full IDE workspace, doesn’t need local tools, or doesn’t require a local Git repo. For example:</p><ul><li>Chatting through ideas out loud with Gemini Live on my phone</li><li>Kicking off some Deep Research, perhaps before persisting in a Gemini Notebook</li><li>Proofreading a LinkedIn post</li><li>Building a quick 3D visualisation in Canvas</li><li>Checking on a recent Google product rebrand (yeah — I have to do this a lot)</li><li>Creating photos and videos</li><li>Comparing flight options</li></ul><h3>A Quick Tour of Pre-Made Connectors in the Gemini App</h3><p>Head to <strong>Settings </strong>→→<strong> Connectors</strong> to open the <strong>Connected apps</strong> hub, where Google groups customisations into <strong>Connectors</strong> and <strong>Skills</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Acrqobpbu8AvysbEe55v3A.png" /><figcaption>Connectors in the Gemini App</figcaption></figure><p>First, take a look at some of the integrations available out of the box:</p><ul><li><strong>Google Workspace:</strong> Read, write, and automate across your Google apps. Draft emails in Gmail, create Google Docs or Sheets, and schedule Calendar events.</li><li><strong>Canva:</strong> Connects to your Canva workspace so you can search your existing Canva folders and designs, generate new presentations or social graphics with your <strong>Brand Kit</strong>, tweak existing templates, resize assets across channels, and export directly to .pptx, .pdf, .png, or .mp4.</li><li><strong>Picsart:</strong> Connects to Picsart’s creative pipeline, unlocking deterministic utilities that raw generative models don’t natively handle — such as true background removal (to .png or .webp), vectorisation for scalable logos and icons, and access to many third-party creative models (like Flux, Ideogram, Recraft, Kling, Runway, and ElevenLabs) right inside your Gemini chat.</li><li><strong>Wispr:</strong> Connect to your Wispr Flow desktop meeting recorder. Let Gemini search your recorded meeting transcripts, scratchpad notes, and action items by participant or date — so you can ask <em>“What did I commit to in yesterday’s architecture sync?”</em> without ever inviting a creepy third-party recording bot into your calls.</li><li><strong>GitHub:</strong> <em>“Import code from public or private repositories, and ask questions about it.”</em> (More on this shortly!)</li></ul><h3>Why I Want GitHub Access in Every Gemini Conversation</h3><p>Everything I build and write in Antigravity lives in GitHub. Not just code and applications, but my article drafts, and presentations too. I keep much of my AI configuration, harness and writing persona in a private GitHub repo.</p><p>So when I’m in the Gemini App or on my phone, I want to say things like:</p><ul><li><em>“Proofread this LinkedIn post using my </em><em>dazbo-content style guide.&quot;</em></li><li><em>“Review this Medium draft against my style guide rules.”</em></li><li><em>“Which public repos have I updated this month?”</em></li><li><em>“How did I structure the ADK evaluation loop in </em><em>harness-engineering-demo?&quot;</em></li></ul><h3>Why the Out-of-the-Box GitHub Connector Is No Good</h3><p>When you see <strong>GitHub</strong> in the <strong>Installed</strong> list, you might assume enabling it gives Gemini tool access to query your GitHub repos on demand. It doesn’t!</p><p>It’s <strong>not</strong> an agentic tool integration. It is a manual <strong>static context importer</strong> attached to the <strong>+</strong> button in the chat bar (<strong>+ </strong>→→<strong> More uploads </strong>→→<strong> Import code</strong>). It has some limitations:</p><ul><li><strong>Manual Per-Chat Uploads:</strong> You must paste a repo URL into a single chat. Start a new chat tomorrow? Import it again.</li><li><strong>Zero Cross-Repo Discovery:</strong> Gemini cannot list your repos or search for files across your account.</li><li><strong>No Live Git Intelligence:</strong> It only ingests a static file snapshot — no commits, PRs, Actions runs, or issues.</li><li><strong>Single-Repo Tunnel Vision:</strong> Mid-conversation, it cannot hop to another repo to check a config file or skill.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sSOCZ3EZa2iSUJHiR_tDBA.png" /><figcaption>GitHub Connector vs Custom</figcaption></figure><h3>Enter the GitHub Remote MCP Server</h3><p>To get on-demand agentic access across all of our repositories in <em>every</em> chat, we can use <strong>GitHub’s managed remote MCP server</strong> (https://api.githubcopilot.com/mcp/) configured as a <strong>Custom Connector</strong>. This server exposes GitHub&#39;s capabilities to any AI agent, such as the Gemini App.</p><p>Better still, GitHub lets us control which toolsets are exposed — and whether Gemini has <strong>read-write</strong> or strictly <strong>read-only</strong> access — simply via the URL path:</p><ul><li><strong>Default Toolset — Read-Only (</strong><strong>https://api.githubcopilot.com/mcp/readonly):</strong> <em>(Recommended)</em> Safe cross-repo exploration, reading private files/skills, code search, and PR/commit analysis with zero mutation risk.</li><li><strong>Default Toolset — Read/Write (</strong><strong>https://api.githubcopilot.com/mcp/):</strong> Workflows where you also want Gemini to create issues, branches, or PRs directly from chat.</li></ul><p>I use https://api.githubcopilot.com/mcp/readonly so Gemini can read <em>anything</em> across my public and private repos without any risk of a stray prompt pushing a commit or closing an issue.</p><h3>How OAuth 2.0 Works Between Gemini and GitHub</h3><p>Because both the Gemini App and the GitHub MCP server live in the cloud, we don’t want to paste our Personal Access Token (PAT) into our chat. Instead, the Gemini App authenticates with GitHub on your behalf using standard <strong>OAuth 2.0</strong>.</p><p>In a typical OAuth setup between an AI client (like Gemini) and a service provider (like GitHub), there are three pieces of configuration:</p><ul><li><strong>Client ID &amp; Client Secret:</strong> Obtained by registering an <strong>OAuth App</strong> in your GitHub Developer Settings. This gives your Gemini connector its own dedicated identity and lets you revoke access at any time with a single click.</li><li><strong>Authorization Callback (Redirect) URI:</strong> The exact URL inside the client application where GitHub is allowed to send you back (along with a temporary authorisation code) after you click <strong>Authorize</strong>. For security, GitHub strictly validates that the redirect URI requested during login matches the callback URL registered on your GitHub OAuth App.</li></ul><p>To set this up, keep <strong>Gemini</strong> and <strong>GitHub</strong> open in two browser tabs: start in the Gemini App to copy the Redirect URI, paste it into GitHub to register your OAuth App, and then copy the resulting Client ID and Secret back into Gemini.</p><h3>Step-by-Step: Adding the Custom GitHub MCP Connector</h3><p>Let’s wire it up. We’ll create a dedicated OAuth App in GitHub specifically for our Gemini connector.</p><h4>Step 1: Open the Custom MCP Modal in Gemini and Grab the Redirect URI</h4><ol><li>Go to <strong>Settings </strong>→→<strong> Connectors</strong> (<a href="https://gemini.google.com/customize/connectors?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">gemini.google.com/customize/connectors</a>) and click <strong>+ Custom</strong>.</li><li>Paste https://api.githubcopilot.com/mcp/readonly into <strong>MCP server URL</strong>.</li><li>Click <strong>Additional settings </strong>→→<strong> Show more</strong>, then click <strong>Copy redirect URI</strong> underneath <strong>Client ID</strong> and <strong>Client secret</strong>.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NE1hAjAEJeAw8oDE0bZLeA.png" /><figcaption>Connect custom MCP Server in Gemini App</figcaption></figure><p>This will generate a unique callback URI bound to your Google User ID:</p><pre>https://oauth-redirect.googleusercontent.com/r/user_bound_custom-mcp-&lt;YOUR_GOOGLE_USER_ID&gt;-api_githubcopilot_com</pre><h4>Step 2: Register Your Dedicated OAuth App in GitHub</h4><ol><li>In GitHub, go to <strong>Settings </strong>→→<strong> Developer settings </strong>→→<strong> OAuth Apps</strong> (<a href="https://github.com/settings/developers">github.com/settings/developers</a>) and click <strong>New OAuth App</strong>.</li><li>Set <strong>Application name</strong> to Gemini MCP Connector and <strong>Homepage URL</strong> to <a href="https://gemini.google.com.">https://gemini.google.com.</a></li><li>Paste the exact https://oauth-redirect.googleusercontent.com/r/user_bound_custom-mcp-... URI into <strong>Authorization callback URL</strong> and click <strong>Register application</strong>.</li><li>Copy your <strong>Client ID</strong>, and click <strong>Generate a new client secret</strong> to copy your <strong>Client Secret</strong>.</li><li>Be sure to <strong>Save</strong> (or Update) the App in GitHub.</li></ol><h4>Step 3: Complete the OAuth Handshake in Gemini</h4><ol><li>Return to the <strong>Connect to an MCP server</strong> dialog in Gemini, paste your <strong>Client ID</strong> and <strong>Client Secret</strong>, and click <strong>Next</strong>.</li><li>Click <strong>Authorize</strong> in the GitHub OAuth popup.</li><li>Accept the privacy policy.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RZWLUobPzuK7hDQjCkyM4w.png" /><figcaption>Custom GitHub MCP Server</figcaption></figure><p>Finally, click <strong>Connect</strong>.</p><p>Once redirected back, <strong>GitHub MCP Server</strong> appears under <strong>Custom apps</strong>:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*abnE6ZRuYSM19LM5MCNQzQ.png" /><figcaption>Custom GitHub MCP Server</figcaption></figure><h3>Testing It Live: Querying Repositories and Loading My Persona</h3><p>In a new Gemini chat, ask a question that needs GitHub access, or tag <strong>@GitHub MCP Server</strong> from the @ menu. For example, when I ran <em>List my public GitHub repos please</em>, Gemini queried the live MCP server and rendered a table of all 33 of my public repositories:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zC5jNVONvTwYc6MMAO8YIA.png" /><figcaption>Asking Gemini App to list my repos</figcaption></figure><p>Then I can ask something like: <em>“Load my </em><em>seo-audit skill from my </em><em>dazbo-config repo.&quot;</em> Gemini loads the skill from my private repo and lets me know we&#39;re ready to use it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CWFrYd9kiB4I72zx9fkZsw.png" /><figcaption>Perform SEO audit using a skill from Gemini App</figcaption></figure><p>Easy peasy!</p><h3>Wrapping Up</h3><p>Antigravity, AI Studio and the Gemini App are all Google AI tools, but they serve different purposes in my daily workflows.</p><p>By wiring <strong>GitHub’s managed remote MCP server</strong> into the Gemini App’s <strong>Custom Connectors</strong> and pairing it with persistent <strong>Gemini Skills</strong>, gemini.google.com evolves from an isolated chat tab into a context-aware workspace. It is connected to my public and private repos, personal docs, and my own style guides and codified persona.</p><p>Give it a go, and let me know in the comments what custom MCP servers and skills you are wiring into your Gemini App!</p><h3>Before You Go</h3><ul><li><strong>Please share</strong> 📢 this with anyone that you think will be interested. It might help them, and it really helps me!</li><li>Please <strong>give me loads of claps</strong>! 👏 (Just hold down the clap button.)</li><li>Please <strong>leave a comment</strong> 💬. Interaction is good!</li><li><strong>Add a star</strong> ⭐ on my repos!</li><li><strong>Follow</strong> 👉 and <strong>subscribe</strong> 🔔, so you don’t miss my content.</li></ul><h3>References and Useful Links</h3><ul><li><a href="https://gemini.google.com/customize/connectors?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Gemini Web App — Connected Apps</a></li><li><a href="https://gemini.google.com/customize/skills?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Gemini Web App — Persistent Skills</a></li><li><a href="https://docs.cloud.google.com/gemini/enterprise/docs/connectors/custom-mcp-server/set-up-custom-mcp-server?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Google Cloud Docs: Set Up a Custom MCP Server Connector</a></li><li><a href="https://docs.github.com/en/copilot/how-tos/provide-context/use-mcp/use-the-github-mcp-server">GitHub Docs: Using the GitHub Remote MCP Server</a></li><li><a href="https://ai.google.dev/gemini-api/docs/ai-studio-quickstart?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Google AI Studio Quickstart</a></li><li><a href="https://ai.google.dev/gemini-api/docs/aistudio-build-mode?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Google AI Studio Build Mode</a></li><li><a href="https://ai.google.dev/gemini-api/docs/logs-datasets?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Google AI Studio — Logs and Datasets</a></li><li><a href="https://antigravity.google/docs/sdk/overview?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Google Antigravity Documentation</a></li><li><a href="https://dev.to/gde/dialling-our-agents-to-11-my-favourite-mcp-servers-3hbm">Dialling Our Agents to 11: My Favourite MCP Servers</a></li><li><a href="https://dev.to/gde/dialling-our-agents-to-11-agent-skills-you-need-to-be-using-2o2k">Dialling Our Agents to 11: Agent Skills You Need to be Using!</a></li><li><a href="https://medium.com/google-cloud/skills-sprawl-when-too-much-of-a-good-thing-bloats-your-ai-agent-167f907e9dc2">Skills Sprawl: When Too Much of a Good Thing Confuses Your AI Agent</a></li><li><a href="https://medium.com/google-cloud/meet-the-antigravity-extension-gemini-code-assist-on-steroids-33367e5a037c">Meet the Antigravity Extension: Gemini Code Assist on Steroids</a></li><li><a href="https://dazbo.co.uk">Dazbo’s Portfolio</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d82c6cca5d85" width="1" height="1" alt=""><hr><p><a href="https://medium.com/google-cloud/supercharging-the-gemini-app-with-custom-mcp-connectors-skills-and-live-github-access-d82c6cca5d85">Supercharging the Gemini App with Custom MCP Connectors, Skills, and Live GitHub Access</a> was originally published in <a href="https://medium.com/google-cloud">Google Cloud - Community</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Spec-Driven Development For-the-Win: Disposable Code, Living Blueprints, and Building Like a Boss]]></title>
            <link>https://medium.com/google-cloud/spec-driven-development-for-the-win-disposable-code-living-blueprints-and-building-like-a-boss-7fe52e7d2d49?source=rss-58a0bd8ba726------2</link>
            <guid isPermaLink="false">https://medium.com/p/7fe52e7d2d49</guid>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[google-gemini]]></category>
            <category><![CDATA[spec-driven-development]]></category>
            <category><![CDATA[harness-engineering]]></category>
            <category><![CDATA[google-antigravity]]></category>
            <dc:creator><![CDATA[Dazbo (Darren Lester)]]></dc:creator>
            <pubDate>Tue, 20 Oct 2026 10:28:53 GMT</pubDate>
            <atom:updated>2026-10-24T07:19:42.622Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Part 3 of the series “Beyond Vibe Coding: The Engineering Blueprint for Reliable AI Agents”</em></p><blockquote>🗺️ The Series Roadmap: Beyond Vibe Coding</blockquote><blockquote>In 2026, software development is drowning in fragmented buzzwords: Harness Engineering, Loop Engineering, Context Engineering, Spec-Driven Development, Evaluators, and Sandboxes. Are they competing frameworks? Redundant prompt techniques? Or something fundamentally new?</blockquote><blockquote>This four-part series unifies these concepts into a single, cohesive architectural blueprint for building software with AI without writing code manually:</blockquote><ul><li><strong><em>Part 1: </em></strong><a href="https://dev.to/gde/from-vibing-chaos-to-reliable-agents-the-what-why-and-how-of-harness-engineering-4ba0"><strong><em>From Vibing Chaos to Reliable Agents: The What, Why, and How of Harness Engineering</em></strong></a><em><br>We diagnose the hangover of vibe coding, unpack the paradigm shift of shifting developer interventions left, establish our bedrock formula (</em><strong><em>Agent = Model + Harness</em></strong><em>), explore the Lewis Hamilton motorsport analogy, and deconstruct the deterministic components that make up a production developer environment.</em></li><li><strong><em>Part 2: </em></strong><a href="https://dev.to/gde/loop-engineering-in-the-harness-autonomous-self-correcting-systems-part-2-155e"><strong><em>Loop Engineering in the Harness: Autonomous, Self-Correcting Systems</em></strong></a><em><br>We fire up the dynamic runtime engine of the harness: deconstructing the core loop components, establishing hybrid evaluation rubrics (deterministic tests + LLM goal fidelity + tool trajectory verification), and applying the Ratchet Principle.</em></li><li><strong><em>Part 3: Spec-Driven Development For-the-Win: Disposable Code, Living Blueprints, and Building Like a Boss (This Article)</em></strong><em><br>We explore Spec-Driven Development (where code is disposable and specifications are the single source of truth), master Google Antigravity’s unified harness and interactive slash commands (</em><em>/grill-me, </em><em>/plan, </em><em>/goal, </em><em>/boost, </em><em>/learn), </em>and combine formal specs on disk with autonomous tactical workflows in your session<em>.</em></li><li><strong><em>Part 4: Scaling the Harness: Practical Loop Patterns, ADK 2 Graphs, Recursive Harness², and Harness Maintenance (Coming Soon)</em></strong><em><br>We bring everything together with practical loop patterns you can wire into any repo (from declarative Markdown and </em><em>loopy to Google ADK 2 graphs), keep context clean across iterations with short-lived subagents, explore Google Research’s recursive Harness</em><strong><em>²</em></strong><em> framework, and cover ongoing harness maintenance.</em></li></ul><h3>Fuel for the Engine: Replacing Vibes with Verified Intent</h3><p>In <a href="https://dev.to/gde/from-vibing-chaos-to-reliable-agents-the-what-why-and-how-of-harness-engineering-4ba0">Part 1</a>, we used the race car analogy for the agentic harness and built the deterministic chassis. In <a href="https://dev.to/gde/loop-engineering-in-the-harness-autonomous-self-correcting-systems-part-2-155e">Part 2</a>, we built the powertrain: the autonomous execution loop, hybrid evaluation rubrics, and the Ratchet Principle.</p><p>Crucially, in Part 2 we established a foundational architectural discipline: <strong>decoupling reusable engineering guardrails from the task specification.</strong></p><p>We centralised the “how we build” — coding standards, test-driven development mandates, security constraints, and automated linting rules — into persistent workspace context (GEMINI.md, AGENTS.md) and Agent Skills. That gave our loop its invariant engineering discipline.</p><p><em>So what fuels the loop? What defines “what we build”?</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xeq8HWoHF1-RC9FonwXFvw.png" /><figcaption>Fueling the Developer Harness</figcaption></figure><p>If you pair rigorous engineering guardrails with vague, casual chat prompts and hand-waving, you won’t get quality software. You’ll get AI slop at 50 tokens a second.</p><p>This is where <strong>Spec-Driven Development (SDD)</strong> enters the frame. As Google Developer Advocate Lee Boonstra brilliantly captured in her whitepaper <a href="https://www.kaggle.com/whitepaper-spec-driven-production-grade-development-in-the-age-of-vibe-coding?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco"><em>Spec-Driven Production Grade Development in the Age of Vibe Coding</em></a>, formalising the task into a structured, testable specification transforms a slop generator into a precision instrument.</p><p>It also unlocks a fundamental architectural shift: <strong>code is disposable, but specifications are the permanent source of truth.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Wgi6QXvZQegYX2hmc5WPZA.png" /><figcaption>The Architectural Shift: Permanent Living Specifications vs Disposable Generated Code</figcaption></figure><p>Because your engineering context, linters, and evaluation rubrics already govern quality, the generated source code is merely an ephemeral byproduct. If business requirements change or an implementation runs down an architectural blind alley, you don’t spend days manually untangling spaghetti code. Your standing guardrails stay untouched; you sharpen the task specification, and let your harness regenerate a clean solution from first principles.</p><p>In <a href="https://arxiv.org/abs/2602.08111"><em>Spec-Driven Development for Agentic Software Engineering: Harnessing Human-Agent Teamwork</em></a>, researchers demonstrated that unconstrained conversational agents inevitably suffer from <strong>compounding specification drift</strong> — a phenomenon where tiny misunderstandings compound across successive turns. Without a fixed specification, it is virtually guaranteed that the final output will drift away from your architectural intent.</p><p>SDD eliminates this drift by introducing immutable contract checkpoints at every phase, keeping the agent firmly on the rails instead of letting it wander off into the weeds.</p><p>For years, the industry swung between two dysfunctional extremes: either spending months drafting bloated, upfront documentation that was obsolete before deployment, or weaponising a poor understanding of “agile” to jump straight into code with zero planning.</p><p>SDD finally closes that gap. Iterating on a half-page specification with an AI partner takes five minutes. The harness then builds and verifies the code in the sandbox in ninety seconds. You inspect working software, get instant empirical feedback, update the contract, and loop. It delivers the deep architectural rigour that enterprises require, at the blistering feedback cadence that agile was always supposed to deliver.</p><p>In other words, <strong>SDD plus an agent takes genuine agile principles to their logical conclusion</strong>.</p><h3>The SDD Lifecycle as a Continuous Loop</h3><p>In an agentic workflow, clear documentation isn’t cumbersome ceremony — <strong>it is a superpower</strong>. Spec-Driven Development is emphatically not a rigid, 200-page document that gathers dust in an enterprise SharePoint graveyard. It is a continuous, living lifecycle that evolves in near real-time alongside the codebase.</p><p>Let’s take a look at the SDD lifecycle:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/688/1*Tmv5euF39g-lfWldsZt1yw.png" /><figcaption>SDD Lifecycle</figcaption></figure><p>Don’t let the formal diagram labels fool you into thinking this is a return to heavyweight waterfall processes. In an agentic workflow, <strong>Steps 1, 2, and 3 happen collaboratively in minutes</strong> — often in a single conversational exchange — to produce that concise living blueprint:</p><ol><li><strong>Scope &amp; Edge Discovery (The WHAT):</strong> A quick dialogue pinning down user intent, non-goals, and boundary conditions. No 50-page PRD; just conversational clarity on what’s in scope and what’s out.</li><li><strong>Interface Contracts &amp; Boundaries (The HOW):</strong> Nailing down the API shape, schema constraints, and Operational Boundaries (<em>Always Do</em>, <em>Ask First</em>, <em>Never Do</em>) before authoring implementation code.</li><li><strong>Acceptance Criteria &amp; Test Scenarios (The PROOF):</strong> Declaring how success will be proven — a few lines of Gherkin or a failing test stub — giving the agent an unambiguous finish line.</li><li><strong>Automated Code Generation:</strong> The agent consumes the living specification within the sandboxed harness, generating implementation code constrained strictly by the contracts and boundaries agreed in steps 1 to 3.</li><li><strong>Testing &amp; Continuous Validation:</strong> Automated test runners execute unit, integration, and static analysis suites directly against the test specification. Simple code defects are patched by Part 2’s inner execution loop, while discovered requirement gaps loop straight back to refine the living spec.</li></ol><h4>Concrete Artifact: A Production Spec Blueprint (tasks/plan.md)</h4><p>What does a real specification look like when fed into an agentic harness?</p><p>It should not be an informal prompt. It is a structured, version-controlled markdown artifact living directly in your repository.</p><p>Here is a <strong>reference blueprint of a production </strong><strong>tasks/plan.md specification</strong> for a cloud service (in this case, an audit event logger backed by Cloud Spanner):</p><pre># Implementation Spec: Cloud Spanner Audit Event Logger<br><br>## 1. Architectural Scope &amp; Contracts<br><br>- **Component:** `AuditEventLogger` service in `services/audit_logger.py`<br>- **Persistence:** Cloud Spanner table `audit_events` (Schema: `event_id STRING(36)`, `timestamp TIMESTAMP`, `actor STRING(128)`, `action STRING(64)`, `payload JSON`)<br>- **Transport:** Async gRPC client via `google-cloud-spanner` Python SDK<br><br>## 2. Operational Boundaries<br><br>### Always Do<br><br>- Enforce strict typing on all method signatures using PEP 585 standard types.<br>- Ensure all database writes use Spanner Mutation batches (`spanner.insert_or_update`) rather than ad-hoc DML statements.<br>- Run `uvx ruff check --fix .` and `uvx codespell -s` after any file modification.<br>- Practice TDD: author failing test in `tests/test_audit_logger.py` before touching service code.<br><br>### Ask First (Human Verification Required)<br><br>- Adding any new third-party Python dependency to `pyproject.toml`.<br>- Modifying Spanner DDL or altering table schemas.<br>- Proposing changes to the shared gRPC authentication interceptor.<br><br>### Never Do<br><br>- Never hardcode service account keys, project IDs, or credentials; rely strictly on Application Default Credentials (ADC).<br>- Never execute inline `python -c` snippets in terminal sessions.<br>- Never catch generic `Exception` without re-raising or logging with structured traceback.<br><br>## 3. Phased Vertical Decomposition<br><br>- [ ] **Phase 1: Domain Models &amp; Test Scaffolding**<br>  - Define `AuditEvent` Pydantic / dataclass model in `models/audit.py`.<br>  - Author unit test mocking Spanner database client in `tests/test_audit_logger.py`.<br>  - *Verification:* `uv run pytest tests/test_audit_logger.py` runs and fails with `NotImplementedError`.<br><br>- [ ] **Phase 2: Service Implementation &amp; Batch Mutations**<br>  - Implement `log_event()` and `log_batch()` methods in `services/audit_logger.py`.<br>  - Implement transient backoff retry logic for `Aborted` and `DeadlineExceeded` Spanner errors.<br>  - *Verification:* `uv run pytest tests/test_audit_logger.py` passes 100%.<br><br>- [ ] **Phase 3: Static Analysis &amp; Hygiene**<br>  - Execute static analysis and spellcheck suite.<br>  - *Verification:* Zero warnings from `ruff` and `codespell`.</pre><p>Notice the power of the <strong>Operational Boundaries</strong> (<em>Always Do</em>, <em>Ask First</em>, <em>Never Do</em>). This is where harness engineering neutralises the wild unpredictability of vibe coding. The agent knows exactly where it has autonomous authority (writing tests, formatting code) and where it must yield control to the human conductor (modifying schemas, adding dependencies).</p><h4>What vs How: The Two Layers of a Living Specification</h4><p>Now, is a single technical plan like our Spanner blueprint the only way to frame a specification? Definitely not.</p><p>In production engineering, specifications often operate across two complementary layers:</p><ol><li><strong>The Functional Specification (</strong><strong>spec.md — The &quot;WHAT&quot; &amp; &quot;WHY&quot;):</strong><br>User-centric and technology-agnostic. It defines the business goals, persona expectations, explicit non-goals (what the agent must <em>not</em> build, preventing scope creep), and verifiable acceptance criteria. It deliberately avoids mentioning specific Python classes, database drivers, or file paths.</li><li><strong>The Technical Implementation Plan (</strong><strong>plan.md — The &quot;HOW&quot;):</strong><br>Codebase-aware and execution-focused. It translates the functional spec into concrete file impact maps (which files to create, edit, or delete), data schemas, library dependencies with explicit versions, and ordered task checklists.</li></ol><p>Our Spanner example above is a <strong>compact technical execution plan</strong> — bundling architectural scope, operational boundaries, and phased vertical slices into a single actionable file for the loop.</p><p>The Operational Boundaries (<em>Always Do</em>, <em>Ask First</em>, <em>Never Do</em>) represent a key operational pattern within that plan: they govern the agent’s blast radius and autonomy envelope. As we’ll explore in the next section, these boundaries can live directly inside a task plan, or be codified globally as a non-negotiable project constitution.</p><h4>Spec Anatomy: Token Physics and Behaviour-Driven Contracts</h4><p>When crafting a specification, two practical disciplines make the difference between an agent that drifts and one that executes with surgical precision:</p><ol><li><strong>Token Physics &amp; Formatting Science (The Hybrid Markdown + YAML Rule):</strong><br>Large language models do not interpret abstract data structures; they process tokenised text. In recent agent compilation research (such as the <em>SkCC</em> study highlighted in Lee Boonstra’s whitepaper), LLM agents suffered up to a <strong>40% performance drop</strong> when fed generic, unoptimised Markdown documents.<br>For Gemini models, the sweet spot is <strong>Hybrid Markdown + Conditional YAML</strong>. Use clean Markdown headers (#, ##) to anchor broad system context and global narrative instructions. Meanwhile, render structured configuration, data contracts, and entity schemas (especially with nesting depth &gt; 3) in <strong>flat, compact YAML</strong> rather than verbose JSON or XML. YAML achieves significantly higher parsing accuracy for deep configurations while eliminating the heavy &quot;reasoning format tax&quot; of JSON closing brackets and quotes.</li><li><strong>Behaviour-Driven Development (BDD) with Gherkin:</strong><br>Vibe prompts describe vague wishes (<em>“make a friendly booking system”</em>). A production-grade spec describes verifiable state transitions. Wrapping acceptance criteria in declarative Gherkin syntax (Given / When / Then) forces the model into a deterministic <strong>State </strong>→→<strong> Action </strong>→→<strong> Outcome</strong> sequence:</li></ol><pre>Scenario: Booking a table with active dietary preferences<br>  Given a customer with dietary preference &quot;gluten-free&quot; recorded in session<br>  When they request a reservation for 2 guests on Friday at 19:00<br>  Then verify table availability via the Cloud SQL reservation tool<br>  And confirm the booking, returning a structured reservation ID</pre><p><em>(Side note: even standard Markdown engines natively support </em><em>gherkin syntax highlighting out of the box — proving just how universally recognised this format already is! Not Medium, of course!)</em></p><p>By providing explicit preconditions, the exact tool to trigger, and the expected outcome, you remove all ambiguity. The agent doesn’t need to guess your intent; it simply fulfils the contract.</p><h3>Practical Tooling for Spec-Driven Development</h3><p>You don’t have to invent your own specification formats or build bespoke SDD tooling from scratch. The agentic engineering community has converged on several robust frameworks, each suited to different developer workflows.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pck6xamdXWxT_DBOcndtsg.png" /><figcaption>SDD frameworks</figcaption></figure><h4>1. The spec-driven-development Skill</h4><ul><li><strong>What it is:</strong> An open-source Agent Skill (created by Addy Osmani) that enforces specification discipline as an autonomous conversational workflow loaded directly into your agent’s context.</li><li><strong>How you use it:</strong> Installed in your workspace (.agents/skills/) or global configuration (~/.gemini/config/skills/), you trigger it conversationally at the start of any feature (e.g. <em>&quot;Use spec-driven development to scaffold the audit logger&quot;</em>). The agent enforces a strict staged sequence (<em>Scope Check → Specify → Plan → Tasks → Implement</em>), pausing at each gate to surface its technical assumptions:</li></ul><pre>ASSUMPTIONS I&#39;M MAKING:<br>1. Target runtime is Python 3.13 managed via uv.<br>2. Data persistence uses Cloud Spanner with dialect GoogleSQL.<br>3. Authentication relies on Google Cloud ADC / IAM principal tokens.<br>-&gt; Confirm or correct before proceeding.</pre><ul><li><strong>What it outputs:</strong> A single, consolidated tasks/plan.md file committed to your repository — exactly like our Spanner blueprint above — featuring architectural contracts, Operational Boundaries (<em>Always Do</em>, <em>Ask First</em>, <em>Never Do</em>), and an ordered checklist of test-verifiable tasks.</li></ul><h4>2. GitHub’s <a href="https://github.com/github/spec-kit">spec-kit</a></h4><ul><li><strong>What it is:</strong> An open-source CLI toolkit and prompt harness created by GitHub that standardises specification-first software development through dedicated commands, project-level constitutions, and multi-file dossiers.</li><li><strong>How you use it:</strong> Executed via CLI or interactive slash commands (/speckit.specify → /speckit.clarify → /speckit.plan → /speckit.tasks → /speckit.implement). You run each step as an explicit command in your terminal or IDE, validating proposed designs against a non-negotiable Project Constitution (.specify/memory/constitution.md). It also includes independent, lightweight workflows for <strong>Bug Fixing</strong> (assess → fix → test) and <strong>Idea Assessment</strong> (intake → define → decide).</li><li><strong>What it outputs:</strong> A modular, version-controlled dossier under specs/&lt;feature-branch&gt;/ containing separated files: spec.md (the user-facing WHAT), plan.md and data-model.md (the technical HOW), research.md (library versions and decisions), and tasks.md (an atomic execution checklist).</li></ul><h4>3. Google’s Conductor Extension</h4><ul><li><strong>What it is:</strong> An AI-native extension and visual workspace assistant from Google that guides developers through context gathering, specification authoring, and test generation directly inside their editor.</li><li><strong>How you use it:</strong> As announced on the <a href="https://developers.googleblog.com/evolving-spec-driven-development-conductor-now-supports-antigravity/?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Google Developers Blog</a>, Conductor now natively supports Antigravity. It lives directly inside your editor sidebar: initiating a structured context-gathering flow, clarifying ambiguous requirements, and validating test coverage before unlocking code generation.</li><li><strong>What it outputs:</strong> An established set of structured, version-controlled Markdown files — specifically product.md (vision and scope), tech-stack.md (dependencies and architecture), spec.md (contracts and acceptance criteria), and plan.md (task execution). Just like any other SDD framework, these files are committed to the repo and loaded straight into Antigravity&#39;s context window as the permanent source of truth.</li></ul><h3>Google Antigravity: The Unified Developer Harness</h3><p>Where does all of this come together in practice?</p><p>Up to this point, we have talked about harness components conceptually: hierarchical context rules, MCP servers, sandboxed environments, evaluation rubrics, and living specifications. But developers need a concrete runtime to orchestrate all these moving parts into a cohesive workflow.</p><p>In the Google ecosystem, that orchestrator is <a href="https://antigravity.google/?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco"><strong>Google Antigravity</strong></a>.</p><p>Antigravity is not another chat assistant bolted into the sidebar of an editor. It is an <strong>AI-native developer harness engine</strong> built from the ground up around the <strong>Plan </strong>→<strong> Do </strong>→<strong> Verify</strong> engineering cycle. Rather than expecting a single model prompt to handle everything in one turn, Antigravity operationalises harness discipline: actively coordinating the engineering environment around the model, managing context, enforcing boundaries, and verifying outcomes.</p><h4>Inside the Harness: What the Engine Coordinates</h4><p>Don’t mistake Antigravity for a passive chat assistant that merely answers questions and suggests snippets. A true developer harness actively coordinates the engineering environment around the model:</p><ul><li><strong>Automatic Rule &amp; Policy Enforcement (No Copy-Pasting Prompts):</strong> Rather than expecting you to copy-paste prompts or manually configure system instructions, Antigravity natively discovers and binds hierarchical rule files (~/.gemini/GEMINI.md, workspace/.agents/rules/, and workspace/GEMINI.md — which is completely interchangeable with workspace/AGENTS.md depending on your team&#39;s conventions). It ensures your coding standards, package policies, and architectural requirements are deterministically bound to every model invocation.</li><li><strong>First-Class Agent Skills &amp; MCP Orchestration:</strong> Antigravity was built from the ground up for progressive disclosure. It tracks installed skills via lightweight metadata headers, dynamically expanding full instruction sets and schemas only when relevant tasks are triggered. Once declared in your configuration, it natively handles tool registration, schema discovery, and execution against both local and remote MCP servers (such as developer-knowledge and Cloud Logging).</li><li><strong>Isolated Sandboxing with Blast-Radius Control:</strong> Commands executed by Antigravity run inside an isolated sandbox by default. The agent has full read/write access to your workspace, but network access and modifications to host-level operating system files require explicit privilege elevation. This gives the agent freedom to compile, test, and refactor code rapidly without placing the developer’s workstation at risk.</li><li><strong>Subagent Spawning &amp; Context Isolation:</strong> Long-running development tasks often require diverse capabilities. When an agent needs to survey documentation across twenty web pages or perform an in-depth accessibility audit, Antigravity can spawn isolated subagents (like dedicated research or browser agents). The subagent executes its bounded task, returns a synthesised result to the parent orchestrator, and terminates — keeping the primary context window lean and focused.</li><li><strong>Subagent-Driven Development (Task Implementers &amp; Reviewers):</strong> Beyond ad-hoc research, structured SDD harnesses split implementation across specialised subagent personas. Rather than cluttering the primary conversation with thousands of lines of intermediate compiler output and syntax churn, Antigravity can dispatch each task to an isolated <strong>Task Implementer</strong> subagent with a pristine context window. Once implementation completes, an independent <strong>Task Reviewer</strong> subagent evaluates the code against the specification contract, passing only a clean verification receipt back to the primary orchestrator.</li></ul><h4>One Core Harness Across Five Developer Surfaces</h4><p>Now, where does this harness live?</p><p>Crucially, <strong>Google Antigravity delivers this identical harness engine across four developer surfaces, alongside a programmatic Python SDK</strong>:</p><ul><li>The <strong>Antigravity App</strong> (formerly Antigravity 2.0) — a dedicated desktop application designed for intent-based building, allowing you to direct autonomous agents, review visual artifacts, and orchestrate projects without having to worry about the underlying code.</li><li>The <strong>Antigravity IDE</strong> — a standalone, AI-first code editor (built on VS Code) for developers who want a complete coding workspace with intelligent tab completions, inline refactoring, and in-editor sidebar assistance.</li><li>The <strong>Antigravity Extension</strong> — brings the complete harness directly into your existing editor of choice (VS Code, JetBrains, Visual Studio, or Zed) so you don’t have to abandon your familiar tools.</li><li>The <strong>Antigravity CLI</strong> — a full Terminal User Interface (TUI) for developers who live in the command line or remote SSH sessions, and want the interactive harness experience in their terminal.</li><li>The <strong>Antigravity SDK</strong> — a Python library for programmatic orchestration, headless automation, scripted workflows, and custom multi-agent systems.</li></ul><p>Across all these surfaces and tools, the underlying configuration engine, context rules (~/.gemini/ and workspace/.agents/), MCP servers, and sandboxed execution runtimes are completely identical.</p><p>This addresses one of the biggest roadblocks in enterprise adoption: <strong>the IDE lock-in barrier</strong>. As I explored in <a href="https://dev.to/gde/meet-the-antigravity-extension-gemini-code-assist-on-steroids-in-whatever-ide-you-love-4p1n"><em>Meet the Antigravity Extension: Gemini Code Assist on Steroids, in Whatever IDE You Love</em></a>, tens of millions of developers — and enterprise security policies — will not simply abandon their favourite IDEs overnight. By packaging the entire unified harness into an extension for VS Code, JetBrains, and beyond, Google ensures teams get full harness intelligence without having to abandon their existing toolchains.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xE80MnRN6zrVZCdHrnfmNw.png" /><figcaption>Antigravity Surfaces and Commands</figcaption></figure><h4>Supercharged Harness Workflows (Slash Commands)</h4><p>Antigravity provides a number of “slash” commands out-of-the-box. Rather than typing out long prompt instructions, you can trigger key harness behaviours directly using these commands:</p><p><strong>/grill-me</strong></p><p>The antidote to underspecified requirements. Instead of accepting a vague prompt and guessing how to implement it, Antigravity initiates a structured, one-question-at-a-time interview to stress-test your thinking, interrogate boundary conditions, and resolve ambiguities <em>before</em> a single file is modified.</p><p><em>Real-world prompt:</em></p><pre>/grill-me on the Spanner schema and backoff retry behaviour in tasks/plan.md</pre><p><strong>/plan</strong></p><p>Switches the agent into dedicated planning mode. Rather than immediately modifying files, Antigravity surveys the codebase, maps out dependencies, and presents a structured implementation plan with verification steps for your review and sign-off before execution begins.</p><p><em>Real-world prompt:</em></p><pre>/plan refactor services/audit_logger.py to use async batch mutations with zero breaking changes</pre><p><strong>/goal</strong></p><p>Initiates an autonomous execution loop for long-running tasks. Rather than pausing after every step to ask for input, the agent works through the task list, diagnoses test failures, handles retries, and yields control only when the objective is fully achieved and verified.</p><p><em>Real-world prompt:</em></p><pre>/goal implement Phase 2 of tasks/plan.md, fix any failing pytest assertions, and ensure ruff passes</pre><p><strong>/boost</strong></p><p>Transforms a single model session into a coordinated team of subagents executing tasks or adversarial reviews in parallel. In loop engineering, /boost is the easiest way to implement <strong>decoupled multi-agent verification</strong> — pairing a builder subagent with an independent critic subagent so a single model instance never marks its own homework.</p><p><em>Real-world prompt:</em></p><pre>/boost run an adversarial security and race-condition review of our Spanner transaction handler</pre><p><strong>/learn</strong></p><p>The mechanical realisation of the Ratchet Principle! When you guide or correct an agent during an interactive session, running /learn extracts the operational lesson and crystallises it directly into your workspace rules. The lesson is immediately persisted to disk so future sessions never repeat the mistake.</p><p><em>Real-world prompt:</em></p><pre>/learn always wrap Spanner Mutation writes in exponential backoff when catching Aborted errors</pre><h4>Slash Commands Alone vs. Formal SDD</h4><p>At this point, I know what you might be thinking:</p><blockquote>“Hold on, Dazbo. If Antigravity already gives me /plan and /goal out of the box, why do I need formal SDD skills like Conductor, spec-kit, or the spec-driven-development skill? Isn&#39;t that just extra ceremony?&quot;</blockquote><p>Great question. Here’s the thing: it comes down to the trade-offs between <strong>tactical session planning</strong> and <strong>long-term architectural governance</strong>:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fsiteproxy.ruqli.workers.dev%3A443%2Fhttps%2Fdatawrapper.dwcdn.net%2FFF2Lo%2F1%2F&amp;display_name=Datawrapper&amp;url=https%3A%2F%2Fsiteproxy.ruqli.workers.dev%3A443%2Fhttps%2Fdatawrapper.dwcdn.net%2FFF2Lo%2F1%2F&amp;image=https%3A%2F%2Fsiteproxy.ruqli.workers.dev%3A443%2Fhttps%2Fdatawrapper.dwcdn.net%2FFF2Lo%2Fplain-s.png%3Fv%3D1&amp;type=text%2Fhtml&amp;schema=dwcdn" width="600" height="653" frameborder="0" scrolling="no"><a href="https://medium.com/media/4b8719eb8b442cca4b53cf842ba70bcc/href">https://medium.com/media/4b8719eb8b442cca4b53cf842ba70bcc/href</a></iframe><h4>An Enterprise Option: <strong>Combining SDD with Tactical Slash Commands</strong></h4><p>You don’t always need both. For everyday feature slices, targeted refactors, or rapid bug fixes, Antigravity’s native slash commands (/plan and /goal) provide plenty of structure with zero setup overhead. Conversely, for architectural reviews, a standalone SDD dossier alone might be all your team requires.</p><p>However, when you are tackling large, multi-component enterprise initiatives, <strong>nesting these layers into a hybrid workflow is a powerful option</strong>. You can use formal SDD as the macro-architectural envelope on disk, and Antigravity’s native slash commands as the micro-execution engine in the session.</p><p>Here is how you can combine them when orchestrating complex features:</p><ol><li><strong>Phase 1 — Macro Architecture (The Outer Spec via Formal SDD / Conductor / </strong><strong>spec-kit):</strong><br>Before generating any implementation code, establish the enduring system blueprint on disk. Define machine-readable API contracts, data schemas, test stubs, and the operational boundary rules (specs/api_contract.md, tasks/plan.md). Because these are committed to version control, they survive across sessions, branches, and team members.</li><li><strong>Phase 2 — Specification Stress-Testing (</strong><strong>/grill-me):</strong><br>Once the initial specification is drafted, invoke /grill-me against the spec file. Antigravity acts as a relentless, sceptical peer reviewer. It interrogates boundary conditions, probes error handling paths, questions concurrency models, and surfaces hidden assumptions <em>before</em> signing off on the architecture.</li><li><strong>Phase 3 — Autonomous Tactical Execution (</strong><strong>/plan + </strong><strong>/goal + </strong><strong>/boost):</strong><br>With the formal spec locked, feed a single vertical slice from your task plan into /goal (or /boost for separated Builder/Critic verification). Antigravity&#39;s autonomous workbench takes over, executing the <strong>tactical inner execution loop we built in Part 2</strong>: it uses /plan to map out immediate file edits, executes changes inside the sandbox, validates against deterministic linters and semantic rubrics, and loops autonomously until every gate is green.</li><li><strong>Phase 4 — Hardening the Harness (</strong><strong>/learn):</strong><br>When the agent encounters an unexpected runtime quirk or discovers a project-specific constraint during the /goal execution loop (such as a subtle gRPC serialisation rule or sandbox path issue), invoke /learn. Antigravity extracts that operational lesson and crystallises it directly into workspace/.agents/rules/ or your root rule file (GEMINI.md / AGENTS.md). The ratchet clicks; the harness hardens; and the entire team permanently benefits from the discovery.</li></ol><h3>Sustainable Team Engineering: Defeating Approval Fatigue</h3><p>Adopting Spec-Driven Development doesn’t just change how you interact with an AI model — it fundamentally alters team engineering culture.</p><p>When teams first adopt automated coding assistants, they often fall into a trap known as <strong>review theatre</strong>. Developers find themselves staring at a 1,200-line diff generated in three seconds, trying to manually verify every line of syntax. Developer experience studies consistently highlight that frequent AI tool users suffer significant cognitive burnout when trapped in manual code reviews. <strong>Eyeball inspection simply does not scale to agentic velocity.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bo9ESD8O2GM3s0vVjAUjow.png" /><figcaption>Review fatigue</figcaption></figure><p>SDD breaks this cycle by establishing a clean division of responsibility:</p><ul><li><strong>Humans review the intent and the contracts:</strong> You evaluate the architectural trade-offs, the boundary constraints, and the acceptance test scenarios.</li><li><strong>The harness verifies the code:</strong> Linters, sandboxed test suites, and evaluation rubrics verify the disposable implementation syntax.</li></ul><p>By shifting human review from syntax nitpicking to spec and contract design, engineering teams eliminate approval fatigue while maintaining rigorous production quality. If the contracts and test scenarios are sound, and the automated verification suite passes in the sandbox, you can merge with confidence.</p><h3>Summary &amp; What’s Next: Scaling the Harness</h3><p>We have now brought together our deterministic chassis (<a href="https://dev.to/gde/from-vibing-chaos-to-reliable-agents-the-what-why-and-how-of-harness-engineering-4ba0">Part 1</a>) and our autonomous loop engine (<a href="https://dev.to/gde/loop-engineering-in-the-harness-autonomous-self-correcting-systems-part-2-155e">Part 2</a>), fuelled by living specifications committed directly to the workspace (<strong>Part 3</strong>) and orchestrated through <strong>Google Antigravity</strong>.</p><p><em>But how do you introduce loop engineering into any existing repository in under five minutes? How do you prevent multi-turn loops from collapsing under context bloat? And what happens when you put a recursive meta-loop around the harness itself?</em></p><p>In <strong>Part 4 (The Finale)</strong>, we take our harness to the next level:</p><ul><li><strong>The Anatomy of an Iteration:</strong> How sandboxed environments, ephemeral subagents, and LIVING_MEMORY.md filter out noise and beat attention dilution.</li><li><strong>Practical Ways to Wire Loops into Any Project (And How to Combine Them):</strong> From zero-code Markdown rules (GEMINI.md + RUBRIC.md) and loopy (LOOPS.md) to /goal and programmatic <strong>Google ADK 2 graphs</strong>.</li><li><strong>Google Research’s </strong>Harness2Harness2<strong> (<em>Harness-Squared</em>):</strong> Recursive agent harnessing (Propose →→ Probe →→ Compose) that optimises the editable components of your harness across tasks without retraining a single model weight.</li><li><strong>Enterprise Production Pillars on Google Cloud:</strong> Bridging your local developer harness to cloud-scale production.</li></ul><p>👉 <strong>Continue to Part 4: Scaling the Harness: Practical Loop Patterns, ADK 2 Graphs, Recursive Harness², and Enterprise Production (Coming Soon)</strong></p><p>See you in the finale!</p><h3>Before You Go</h3><ul><li><strong>Please share</strong> 📢 this with anyone that you think will be interested. It might help them, and it really helps me!</li><li>Please <strong>give me loads of claps</strong>! 👏 (Just hold down the clap button.)</li><li>Please <strong>leave a comment</strong> 💬. Interaction is good!</li><li><strong>Add a star</strong> ⭐ on my repos!</li><li><strong>Follow</strong> 👉 and <strong>subscribe</strong> 🔔, so you don’t miss my content.</li></ul><h3>References and Useful Links</h3><h4>Spec-Driven Development</h4><ul><li><a href="https://codelabs.developers.google.com/sdd-adk-antigravity?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Spec-Driven ADK Agent Development with Antigravity and Spec-kit | Google Codelabs</a></li><li><a href="https://codelabs.developers.google.com/codelabs/getting-started-with-spec-driven-development-in-antigravity?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Getting started with Spec Driven Development in Antigravity | Google Codelabs</a></li><li><a href="https://codelabs.developers.google.com/sdd-agy-cli?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Spec-Driven Development with Antigravity CLI — Structured Agent Workflows with Skills and MCP | Google Codelabs</a></li><li><a href="https://developers.googleblog.com/evolving-spec-driven-development-conductor-now-supports-antigravity/?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco"><em>Evolving Spec-Driven Development: Conductor Now Supports Antigravity</em> | Google Developers Blog</a></li><li><a href="https://www.kaggle.com/whitepaper-spec-driven-production-grade-development-in-the-age-of-vibe-coding?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco"><em>Spec-Driven Production Grade Development in the Age of Vibe Coding</em> by Lee Boonstra (Google)</a></li><li><a href="https://arxiv.org/abs/2602.08111"><em>Spec-Driven Development for Agentic Software Engineering: Harnessing Human-Agent Teamwork</em></a></li><li><a href="https://github.com/github/spec-kit">github/spec-kit: Toolkit to help you get started with Spec-Driven Development</a></li><li><a href="https://github.com/addyosmani/agent-skills/tree/main/skills/spec-driven-development">spec-driven-development skill by Addy Osmani</a></li></ul><h4>Google Antigravity &amp; Ecosystem</h4><ul><li><a href="https://antigravity.google/docs/home/?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Google Antigravity Documentation</a></li><li><a href="https://antigravity.google/blog/introducing-google-antigravity?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Introducing Google Antigravity | Google Antigravity Blog</a></li><li><a href="https://medium.com/google-cloud/the-ai-native-developer-experience-part-2-harness-engineering-with-google-antigravity-7fb72dab243f"><em>The AI-Native Developer Experience, Part 2: Harness Engineering with Google Antigravity</em> by Jaroslav Pantsjoha</a></li></ul><h4>Dazbo’s Related Articles</h4><ul><li><a href="https://dev.to/gde/from-vibing-chaos-to-reliable-agents-the-what-why-and-how-of-harness-engineering-4ba0"><em>From Vibing Chaos to Reliable Agents: The What, Why, and How of Harness Engineering (Part 1)</em> by Dazbo</a></li><li><a href="https://dev.to/gde/loop-engineering-in-the-harness-autonomous-self-correcting-systems-part-2-155e"><em>Loop Engineering in the Harness: Autonomous, Self-Correcting Systems (Part 2)</em> by Dazbo</a></li><li><a href="https://dev.to/gde/meet-the-antigravity-extension-gemini-code-assist-on-steroids-in-whatever-ide-you-love-4p1n"><em>Meet the Antigravity Extension: Gemini Code Assist on Steroids, in Whatever IDE You Love</em> by Dazbo</a></li><li><a href="https://dev.to/gde/dialling-our-agents-to-11-agent-skills-you-need-to-be-using-2o2k"><em>Dialling Our Agents to 11: Agent Skills You Need to be Using!</em> by Dazbo</a></li><li><a href="https://dazbo.co.uk"><em>Dazbo’s Portfolio</em></a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7fe52e7d2d49" width="1" height="1" alt=""><hr><p><a href="https://medium.com/google-cloud/spec-driven-development-for-the-win-disposable-code-living-blueprints-and-building-like-a-boss-7fe52e7d2d49">Spec-Driven Development For-the-Win: Disposable Code, Living Blueprints, and Building Like a Boss</a> was originally published in <a href="https://medium.com/google-cloud">Google Cloud - Community</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Automated GitHub Code Reviews Using Google Gemini Update: Enterprise-Ready, Multimodal, and…]]></title>
            <link>https://medium.com/google-cloud/automated-github-code-reviews-using-google-gemini-update-enterprise-ready-multimodal-and-e3459f621aea?source=rss-58a0bd8ba726------2</link>
            <guid isPermaLink="false">https://medium.com/p/e3459f621aea</guid>
            <category><![CDATA[ci-cd-pipeline]]></category>
            <category><![CDATA[pull-request]]></category>
            <category><![CDATA[code-review]]></category>
            <category><![CDATA[google-gemini]]></category>
            <category><![CDATA[github-actions]]></category>
            <dc:creator><![CDATA[Dazbo (Darren Lester)]]></dc:creator>
            <pubDate>Mon, 19 Oct 2026 05:34:22 GMT</pubDate>
            <atom:updated>2026-10-19T08:22:36.113Z</atom:updated>
            <content:encoded><![CDATA[<h3>Automated GitHub Code Reviews Using Google Gemini Update: Enterprise-Ready, Multimodal, and Battle-Hardened</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*83nbCFu4tKRHtQTt6ojnBw.png" /><figcaption>Gemini PR Review &amp; Triage</figcaption></figure><p>You are in <strong>Part 2</strong> of the series <strong><em>&quot;Automated GitHub Code Reviews Using Google Gemini&quot;</em></strong></p><blockquote><em>🗺️ Series Navigation</em></blockquote><blockquote><strong>Part 1:</strong> <a href="https://dev.to/gde/automated-github-code-reviews-using-google-gemini-4l0f"><strong>Automated GitHub Code Reviews Using Google Gemini</strong></a> — Why I built a super-fast, native composite GitHub Action to replace Google&#39;s sunsetted run-gemini-cli and Gemini Code Assist integrations, and how to get up and running in three minutes.</blockquote><blockquote><strong>Part 2: Enterprise-Ready AI Code Reviews &amp; Issue Triage with Google Gemini (This Article)</strong> — How the action evolved from a fast weekend replacement into an enterprise-grade, multimodal code review and triage engine — featuring Gemini 3.8 Flash, visual PR diffing, bounded dynamic context selection, Gemini Enterprise Agent Platform billing attribution, supply-chain hardening, and battle-tested production fixes.</blockquote><h3>Back for the Update!</h3><p>Howdy friends, and welcome back!</p><p>When I published <a href="https://dev.to/gde/automated-github-code-reviews-using-google-gemini-4l0f">Part 1</a> back in July, I had a very specific itch to scratch. Google had just turned off the lights on both run-gemini-cli and the consumer Gemini Code Assist GitHub integration as part of the transition to the <a href="https://antigravity.google/?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Google Antigravity suite</a>. And consequently, my automated agentic pull request reviews stopped working. Sad times.</p><p>So, I rolled up my sleeves...</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*G5hEx5Tp82JT7_HoRBp3Ug.png" /><figcaption>Sleeves rolled up</figcaption></figure><p><em>(Okay, that&#39;s actually an attempt at using AI to put me in a Don Johnson outfit.)</em></p><p>... And built a drop-in replacement: the <a href="https://github.com/derailed-dash/gemini-review-action"><strong>Gemini PR Review &amp; Triage Action</strong></a> (also available directly on the <a href="https://github.com/marketplace/actions/dazbo-s-gemini-code-review-issue-triage">GitHub Marketplace</a>).</p><p>Since launching Part 1, something awesome happened. Not only did I start dogfooding the action heavily across every single repository I maintain, but developers in the community started adopting it — and <strong>it is now actively running in production inside at least one enterprise organisation</strong>!</p><p>A massive shoutout here goes to my fellow Google Developer Expert (GDE), <a href="https://github.com/SaschaHeyer"><strong>Sascha Heyer</strong></a>. Sascha rolled the action out across real production repositories, battle-tested it against enterprise edge cases, and contributed a phenomenal series of pull requests — from real-dollar cost estimation and Gemini Enterprise Agent Platform billing labels to automated review thread resolution, per-file content caps, competing bot filtering, and supply-chain lockfile hardening.</p><p>Through dozens of releases (taking us from v1.4 all the way to v1.8.3), this collaboration has transformed the tool from a &quot;fast drop-in replacement&quot; into what I genuinely believe is the <strong>de facto Google Gemini-based code reviewer and issue triager for GitHub</strong>.</p><p>Even better: it is <strong>100% free and open-source</strong>. There are no SaaS subscriptions and no per-seat licensing fees. You simply bring your own Gemini API key or Google Cloud project, and you pay only for the raw tokens you consume — which, thanks to our built-in context caching, typically works out to <strong>$0.02 to $0.04 per review</strong>!</p><p>In this post, I&#39;m going to:</p><ol><li><strong>Recap the What, the Why, and the Unique Selling Points (USPs)</strong> for anyone new to the party.</li><li><strong>Share the complete current feature list</strong> so you can see everything the action packs under the bonnet today.</li><li><strong>Deep-dive into the major new features, architectural upgrades, and battle-hardened bug fixes</strong> we&#39;ve shipped since Part 1 — including why this solution is now thoroughly <strong>enterprise-ready</strong>.</li></ol><p>Grab a piña colada (or a strong coffee), and let&#39;s dive in!</p><h3>Recap: The What, The Why, and The USPs</h3><p>If you missed <a href="https://dev.to/gde/automated-github-code-reviews-using-google-gemini-4l0f">Part 1</a>, here is the sixty-second elevator pitch on why automated AI code reviews in CI/CD are such a game-changer — and why this specific action hits different.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pLD2hCzMxZN3gseMVtTp6A.png" /><figcaption>Automated Code Review Pipeline with Gemini AI</figcaption></figure><h4>The Problem with Traditional (and Local AI) Code Reviews</h4><p>In a traditional workflow, you open a Pull Request and wait hours (or days) for a human colleague to context-switch away from their own work and review your code. When they finally get to it, human reviews are notoriously variable: some reviewers rubber-stamp a 500-line diff with &quot;LGTM&quot;, while others bikeshed variable names and miss a glaring SQL injection vulnerability.</p><p><em>&quot;Ah,&quot;</em> you might say, <em>&quot;but I already use an AI coding assistant in my IDE!&quot;</em></p><p>Me too! I live in Google Antigravity every day. But relying solely on your local IDE agent to review the code it just helped you write introduces a massive blind spot: <strong>Session Bias</strong>.</p><h4>1. The &quot;Clean Slate&quot; Advantage</h4><p>When you pair-programme with an AI assistant in your IDE, that agent accumulates your entire conversational history. It knows what you tried, why you took a shortcut, and what compromises you made to get the tests passing. Because it understands your <em>intent</em> so well, it frequently forgives subtle regressions, missing edge cases, or architectural drift.</p><p>By running our Gemini code reviewer inside a <strong>GitHub Action on every Pull Request</strong>, the reviewer starts with a <strong>completely clean slate</strong>. It has zero knowledge of your local chat history or how much you struggled with a module. It objectively evaluates the raw git diff against your repository&#39;s codebase context and coding standards. Time and time again, this stateless reviewer catches bugs, security gaps, and documentation discrepancies that my local IDE session glossed right over!</p><h4>2. Blazing Speed: Native Composite Action + Astral&#39;s uv</h4><p>Legacy actions (like the old run-gemini-cli) built or pulled hefty Docker containers on every single run, or used sluggish pip install steps that burned through GitHub Actions runner minutes before a single token was ever sent to Gemini.</p><p>I architected gemini-review-action as a <strong>native Composite Action</strong> powered by Astral&#39;s blazingly fast <strong>uv</strong> package manager. There are no Docker container builds and no slow dependency resolution phases. The environment boots in a couple of seconds and posts line-specific review feedback almost immediately.</p><h4>3. Deep Context + 90% Cost Savings via Gemini Context Caching</h4><p>Most basic AI review scripts only feed the raw git diff to the model. Without knowing what the rest of your repository looks like, an LLM can&#39;t tell if a change breaks an interface defined in another module or violates your project&#39;s architectural patterns.</p><p>Our action features a <strong>Hybrid Codebase Context Engine</strong> paired with <strong>Stateless Gemini Context Caching</strong>:</p><ul><li>For repositories under 1.5 MB of text, it loads the <strong>entire codebase</strong> into context (<em>Full Context Mode</em>).</li><li>For massive repositories and monorepos over 1.5 MB, it switches to <em>Sparse Context Mode</em>, feeding the full directory tree, core manifests/documentation (README.md, ARCHITECTURE.md, AGENTS.md, pyproject.toml, etc.), and dynamically selecting the most relevant sister modules.</li><li>Whenever repository context exceeds ~32k tokens (about 100K characters), the action provisions an ephemeral, server-side <strong>Gemini Context Cache</strong> (CachedContent). Subsequent PR pushes and multi-turn tool calls reuse that cache handle automatically — delivering a <strong>90% discount on cached input tokens</strong> without requiring any local runner disk persistence. <em>(Enterprise DLP Note: On Gemini Enterprise Agent Platform, the cache is strictly tenant-isolated within your Google Cloud project and region, bounded by an automatic 1-hour TTL, and fully covered by Google Cloud&#39;s zero-training enterprise data governance commitments.)</em></li></ul><h3>The Complete Feature Matrix (What&#39;s in the Box Today)</h3><p>Before we unpack the new engineering deep-dives, here is the complete summary of what <strong>Dazbo&#39;s Gemini PR Review &amp; Triage Action (currently at version </strong><strong>1.8.3)</strong> gives you right out of the box:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Da8SJCPAi-LaD0L2XTmlHQ.png" /><figcaption>Gemini PR Review &amp; Triage Features Overview</figcaption></figure><h4>Core Review &amp; Triage Capabilities</h4><ul><li><strong>AI-Powered Pull Request Reviews</strong>: Constructive, severity-graded (🔴 Critical, 🟠 High, 🟡 Medium, 🟢 Low), line-specific inline feedback powered by the latest Gemini Flash model (currently <strong>gemini-3.8-flash</strong>) by default.</li><li><strong>One-Click Interactive Suggestions</strong>: Formats exact code fixes inside native GitHub ```suggestion blocks so developers can commit fixes directly in the GitHub UI with a single click.</li><li><strong>Multi-Axis Quality Framework</strong>: Evaluates every PR systematically across five core engineering axes: <em>Correctness &amp; Verification</em>, <em>Readability &amp; Simplicity</em>, <em>Architecture &amp; System Design</em>, <em>Security &amp; Data Hardening</em>, and <em>Performance &amp; Efficiency</em>.</li><li><strong>Holistic Documentation &amp; Blog Review</strong>: Evaluates Markdown documentation, READMEs, ADRs, and technical articles alongside source code for accuracy and completeness.</li><li><strong>Automated Issue Triage</strong>: Dynamically classifies, prioritises, and applies repository labels to incoming or reopened GitHub issues based on issue descriptions, discussion comments, and attached screenshots.</li><li><strong>On-Demand </strong><strong>/gemini-review Trigger</strong>: Re-trigger reviews at any time via PR comments, protected by strict OWNER, MEMBER, and COLLABORATOR authorisation guards so random internet strangers can&#39;t drain your API quota.</li></ul><h4>Multimodal &amp; Context Intelligence</h4><ul><li><strong>Multimodal Context &amp; Visual PR Diffing</strong>: Automatically extracts local images and PDFs referenced in Markdown files, and performs <strong>side-by-side visual PR diffing</strong> (base_sha vs head_sha) whenever images (.png, .jpg, .webp, .gif, .svg) are modified in a PR.</li><li><strong>Multimodal Issue Triage (Remote Screenshots)</strong>: Automatically extracts and securely fetches remote screenshots embedded in GitHub issues (including private repo user-attachments) so Gemini can visually inspect UI bug reports and stack-trace screenshots.</li><li><strong>Deterministic Image Downscaling</strong>: Automatically resizes large images (&gt;600 KB) down to a maximum of 300 KB before sending to Gemini, protecting your token budget and keeping latency low while passing SVGs and PDFs through lossless.</li><li><strong>Hybrid Codebase Context &amp; Bounded Dynamic Selection</strong>: Seamless switching between Full Context Mode and Sparse Context Mode, using proximity/import heuristics and a 4-tier architectural selector to pick the best sister files in large monorepos.</li><li><strong>Per-File Content Truncation &amp; Token Budget Safeguards</strong>: Caps individual massive files (like generated OpenAPI schemas) at 128 KB and runs preflight token checks so a single huge file never crashes your review.</li><li><strong>Google Developer Knowledge MCP Integration</strong>: Queries official, live Google documentation (for Google Cloud, Gemini, Firebase, Android, etc.) via Model Context Protocol (MCP) during the review (automatically activated when the Developer Knowledge API is enabled on your API key or GCP project).</li><li><strong>On-Demand Agent Skills Registry</strong>: Dynamically discovers and loads relevant coding standards from .agents/skills/. And a curated set of starter skills for Google Cloud, ADK, and Gemini APIs are already bundled.</li></ul><h4>Conversational Memory &amp; Workflow Automation</h4><ul><li><strong>PR Comment &amp; Discussion Thread History</strong>: Retrieves prior inline review threads and PR comments across commits. Follows an opinionated resolution matrix: celebrates resolved items (### ✅ Resolved Items from Prior Reviews), respects developer deferrals and architectural justifications, and re-flags silently ignored issues.</li><li><strong>Auto-Resolving Addressed GitHub Threads</strong>: Optionally resolves GitHub review conversation threads automatically (resolve_addressed_threads: &#39;true&#39;) once Gemini verifies the fix, unblocking <em>&quot;Require conversation resolution before merging&quot;</em> branch protection rules.</li><li><strong>Competing Bot Author Exclusion</strong>: Filter out comments from secondary bots (exclude_comment_authors: &#39;claude[bot]&#39;) so your Gemini review remains genuinely independent and you don&#39;t pay input tokens to process another bot&#39;s prose.</li><li><strong>Smart Inline-Suggestion Commit Skipping</strong>: Automatically detects commits generated by clicking &quot;Commit suggestion&quot; in the GitHub UI and skips redundant re-reviews (skip_inline_suggestions: &#39;true&#39;).</li></ul><h3>Customisation, FinOps &amp; Enterprise Security</h3><ul><li><strong>Zero-Config Agentic Installation Skill</strong>: Equip your IDE agent via npx skills add derailed-dash/gemini-review-action -y -g --skill install-gemini-code-review-action, then simply ask it to <em>&quot;Set up Gemini review in this repo&quot;</em> — your agent handles configuration preferences, template generation, and secret setup guidance automatically.</li><li><strong>Extending vs Replacing Guardrails</strong>: Easily append organisation- or team-specific rules via .github/review-instruction-additions.md or the custom_instructions workflow input without losing the built-in quality framework — or you can override the entire prompt via .github/commands/gemini-review.toml.</li><li><strong>Configurable Personas &amp; Locales</strong>: Choose your reviewer&#39;s language (English (UK) by default) and personality (straight, dazbo, palpatine, or rick). Careful now... Palpatine is <em>not very forgiving.</em></li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0FiLl2SinihXPv70K1F-yQ.png" /><figcaption>The Emperor is not as forgiving as I am</figcaption></figure><ul><li><strong>Reasoning &amp; Thinking Budget Control</strong>: Defaults to <strong>dynamic thinking</strong> (the recommended out-of-the-box setting for PR reviews, letting Gemini scale reasoning to match diff complexity). You can also clamp thinking_level (low, medium, high, or an exact integer token budget) for predictable FinOps cost ceilings, or set it to &#39;off&#39; to make Issue Triage blazing fast and ultra-cheap.</li><li><strong>Tokenomics &amp; Dollar Cost Telemetry</strong>: Appends a collapsible table to every review breaking down uncached input, cached input, comment history, thinking tokens, output tokens, and the estimated dollar cost.</li><li><strong>Enterprise Keyless Authentication (WIF)</strong>: Full support for Google Cloud <strong>Workload Identity Federation (WIF)</strong> and Application Default Credentials (ADC) via Gemini Enterprise Agent Platform (formerly Vertex AI) — zero long-lived API keys required!</li><li><strong>Gemini Enterprise Agent Platform Billing Labels</strong>: Automatically tags Enterprise Agent Platform requests with component and repo (plus custom billing_labels like team=platform,cost_centre=engineering) for granular FinOps chargeback in BigQuery billing exports.</li><li><strong>Supply-Chain &amp; Runtime Hardening</strong>: Enforces immutable dependency lockfiles (uv run --locked), SHA-pinned internal actions, path-traversal containment, and automatic fork PR skipping.</li></ul><h3>Enterprise-Ready by Design: Security, Governance, and FinOps</h3><p>When hobbyists pick a GitHub Action, they just want to paste an API key into a secret and see comments appear on their PR. When enterprise platform and security teams evaluate a GitHub Action, they ask a very different set of questions:</p><ul><li><em>&quot;How do we authenticate without storing static, long-lived API keys across 200 repositories?&quot;</em></li><li><em>&quot;How do we attribute AI spend back to specific business units, teams, and repositories in our Cloud Billing export?&quot;</em></li><li><em>&quot;What stops a malicious dependency drift or an untrusted fork PR from exfiltrating credentials?&quot;</em></li></ul><p>These aren&#39;t theoretical questions — this reviewer is already deployed in production at an enterprise organisation, and many of the capabilities below were forged directly from real-world enterprise governance requirements. Let&#39;s look at the key pillars.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*P7Ayv99uMa23zGwHqR76Gg.png" /><figcaption>Enterprise Governance: Keyless WIF &amp; FinOps Attribution</figcaption></figure><h4>1. Keyless Enterprise Auth with Workload Identity Federation (WIF)</h4><p>Static API keys (GEMINI_API_KEY) are great for quick personal setups. But in an enterprise environment, long-lived credentials stored in GitHub Secrets are a compliance headache. Somebody has to rotate them, audit them, and worry about them leaking.</p><p>Because our action is built on the modern google-genai SDK, it natively supports <strong>Google Cloud Workload Identity Federation (WIF)</strong> with <strong>Application Default Credentials (ADC)</strong> with Gemini Enterprise Agent Platform (formerly Vertex AI).</p><p>Instead of generating a static key, your GitHub Actions runner requests a short-lived OIDC token from GitHub and exchanges it via google-github-actions/auth for an ephemeral Google Cloud access token scoped to a specific Service Account:</p><pre>jobs:<br>  review:<br>    runs-on: ubuntu-latest<br>    permissions:<br>      contents: read<br>      pull-requests: write<br>      id-token: write # Required for WIF OIDC token exchange<br><br>    steps:<br>      - name: Checkout repository<br>        uses: actions/checkout@v6<br>        with:<br>          ref: ${{ github.event.pull_request.head.sha || format(&#39;refs/pull/{0}/head&#39;, github.event.issue.number) }}<br><br>      - name: Authenticate to Google Cloud via WIF<br>        uses: google-github-actions/auth@v2<br>        with:<br>          workload_identity_provider: &#39;projects/123456789012/locations/global/workloadIdentityPools/${{ secrets.WIF_POOL_ID }}/providers/${{ secrets.WIF_PROVIDER_ID }}&#39;<br>          service_account: &#39;${{ secrets.GCP_SERVICE_ACCOUNT }}&#39;<br><br>      - name: Run Gemini Review Action<br>        uses: derailed-dash/gemini-review-action@v1<br>        env:<br>          GOOGLE_GENAI_USE_VERTEXAI: &quot;True&quot;<br>          GOOGLE_CLOUD_PROJECT: &quot;my-enterprise-project-id&quot;<br>          GOOGLE_CLOUD_LOCATION: &quot;global&quot;<br>        with:<br>          github_token: ${{ secrets.GITHUB_TOKEN }}<br>          gemini_model: &#39;gemini-3.8-flash&#39;</pre><p><strong>Why this matters for the enterprise:</strong></p><ul><li><strong>Zero static secrets</strong>: Credentials live for minutes and expire automatically when the job finishes.</li><li><strong>IAM &amp; VPC-SC Governance</strong>: All Gemini API calls route through your Google Cloud project, meaning they respect your organisation&#39;s IAM policies, data residency regions, Cloud Audit Logs, and VPC Service Controls.</li></ul><h4>2. FinOps Transparency: Real Dollar Cost Estimation &amp; Gemini Enterprise Agent Platform Billing Labels</h4><p><strong>ROI is paramount.</strong> If a solution costs too much or you can&#39;t measure what it costs, it&#39;s bad architecture.</p><p>In Part 1, our telemetry block reported raw token counts. That was useful, but as fellow GDE <strong>Sascha Heyer</strong> rightly pointed out when he contributed PR #33, developers and engineering leaders don&#39;t pay their cloud bills in tokens — they pay in dollars!</p><p>We upgraded the telemetry engine (gemini_review/pricing.py) to calculate and print the <strong>estimated dollar cost</strong> on every single review, broken down by uncached input, cached input (at the 90% discount rate), thinking/reasoning tokens, and response output tokens:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/752/1*k5dlyH___NFRInVTmWDKLA.png" /><figcaption>Automated PR Review Output with Dollar Cost Telemetry</figcaption></figure><p>Notice a couple of deliberate engineering choices here:</p><ul><li><strong>No Silent Guessing</strong>: If you configure an experimental or future model that isn&#39;t in our pricing table yet, the action reports the token counts and omits the dollar figure rather than borrowing another model&#39;s rate. A missing number is obvious; a wrong number is dangerous.</li><li><strong>Time-Boxed Introductory Rates</strong>: Models like gemini-3.8-flash have promotional introductory pricing ($0.75 / $3.75 per 1M tokens through 2026-12-31, reverting to $1.50 / $7.50 afterwards). We encoded the promotional expiry dates directly into the pricing engine data structures, and the review footnote explicitly states which rate was applied. And if your enterprise has negotiated custom flat rates or uses Batch/Flex tiers, you can override the table via GEMINI_RATE_INPUT and GEMINI_RATE_OUTPUT.</li></ul><h4>Answering the FinOps Question at Scale: Cloud Billing Labels (billing_labels)</h4><p>Knowing that PR #42 cost $0.03 is great for a developer. But a Head of Platform Engineering needs to answer a different question: <strong>&quot;What did AI code reviews cost across our 150 repositories this month, broken down by engineering team?&quot;</strong></p><p>You shouldn&#39;t have to scrape GitHub PR comments to answer a FinOps question. So we added native <strong>Gemini Enterprise Agent Platform Billing Labels</strong>. Whenever you run on Gemini Enterprise Agent Platform (Vertex AI), the action automatically attaches sanitised GCP billing labels to every generate_content request:</p><ul><li>component: gemini-review-action</li><li>repo: owner_repo</li></ul><pre>  - name: Run Gemini Review Action<br>    uses: derailed-dash/gemini-review-action@v1<br>    with:<br>      github_token: ${{ secrets.GITHUB_TOKEN }}<br>      billing_labels: &#39;team=platform,cost_centre=engineering,env=ci&#39;</pre><p><em>(Pro-Tip: We deliberately did </em><strong><em>not</em></strong><em> make the PR number a default label, because high-cardinality labels turn your BigQuery billing export into a noisy mess. And if you’re using a Google AI Studio API key instead of Gemini Enterprise Agent Platform, the action cleanly omits the </em><em>labels parameter.)</em></p><h4>3. Supply-Chain Security &amp; Fork PR Hardening</h4><p>A code review action holds pull-requests: write permissions and runs against untrusted PR branches. If you aren&#39;t careful, that&#39;s a prime target for supply-chain attacks. Here is how we locked it down:</p><ol><li><strong>Enforced Lockfiles (</strong><strong>uv run --locked)</strong>: Previously, running uv run --project would silently resolve and rewrite uv.lock at runtime if pyproject.toml drifted. That meant even if an enterprise consumer pinned our action to an immutable 40-character Git commit SHA, a transitive PyPI dependency could theoretically change at runtime. So we enforced uv run --locked and pinned all nested setup actions (actions/setup-python and astral-sh/setup-uv) to exact commit SHAs.</li><li><strong>Untrusted Fork PR Safeguards</strong>: By default, GitHub strips repository secrets from fork PRs to prevent untrusted contributors from accessing API keys. If an action tries to run automatically on a fork without secrets, it crashes with an authentication error and paints the contributor’s PR with an ugly red ❌. Our starter workflows solve this gracefully:<br>- <strong>Clean skips on forks</strong>: The workflow guards automatic runs (github.event.pull_request.head.repo.full_name == github.repository), quietly skipping external forks so CI status checks stay green.<br>- <strong>Maintainer on-demand trigger</strong>: Once a repository maintainer (OWNER, MEMBER, COLLABORATOR) inspects the fork code, they can safely trigger Gemini by typing /gemini-review in a PR comment — which runs under their authenticated context with full access to secrets.</li><li><strong>Path Traversal Containment</strong>: Every file path, Markdown image link, and skill reference loaded by the action is strictly validated against the workspace root. An attacker cannot craft a malicious Markdown link like ![diagram](../../../../etc/some-file) or a custom instruction path to trick the runner into reading files outside the repository.</li></ol><h3>Major Feature Deep-Dives Since Part 1</h3><p>Now let’s get into the really fun stuff! Here are the major architectural features we’ve engineered since publishing Part 1.</p><h4>1. Multimodal Context Inspection, Visual PR Diffing &amp; Screenshot Triage</h4><p>Until recently, AI code reviewers had a massive blind spot: <strong>they were completely blind to anything that wasn’t plain text.</strong></p><p>If you updated an architecture diagram in docs/architecture.png, tweaked a UI component screenshot, attached a PDF specification, or opened a GitHub Issue with a screenshot of a broken button, a text-only reviewer saw nothing.</p><p>In our solution, we’ve unlocked Gemini’s native multimodal superpowers across three distinct workflows:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*89kb6BJYFY_dxDqSC3J6gg.png" /><figcaption>Multimodal Context &amp; Visual PR Diffing Engine</figcaption></figure><h4>A. Automatic Markdown Image &amp; PDF Resolution</h4><p>Whenever a Markdown file (README.md, DESIGN.md, ARCHITECTURE.md, or a blog post!) is included in the review context or modified in a PR diff, we scan the document for standard Markdown and HTML image references. We resolve the paths relative to the Markdown file, verify containment inside the repo, and attach the visual assets directly to the Gemini request. Now, when Gemini reviews your documentation updates, it actually <strong>looks at your architecture diagrams</strong> to verify that the diagram matches your code and prose!</p><h4>B. Side-by-Side Visual PR Diffing</h4><p>What happens when a PR modifies an existing architecture diagram, UI component snapshot, or application icon?</p><p>Most AI reviewers either ignore modified images completely or summarise them in separate, isolated passes — which completely loses subtle visual regressions.</p><p>Instead, our action retrieves <strong>both the “before” and “after” versions</strong> of any modified graphic and passes them side-by-side into a single Gemini turn. Because both images sit together in the same multimodal attention window, Gemini performs direct visual diffing — spotting accidentally cropped text, resolution degradation, broken UI alignment, or missing arrows in an updated architecture diagram!</p><h4>C. Multimodal Issue Triage for Remote Screenshots</h4><p>When users report bugs in GitHub Issues, they love pasting screenshots directly into the issue editor (which GitHub hosts at extensionless URLs like https://github.com/user-attachments/assets/&lt;uuid&gt;).</p><p>We upgraded our triage to extract remote HTTPS image URLs from both the issue body and discussion comments:</p><ul><li><strong>Safe Token Handling</strong>: Images in private repositories require GitHub credentials to view. The downloader attaches your GITHUB_TOKEN only to trusted GitHub domains — and strips it immediately if a link redirects to an external site.</li><li><strong>Attack-Proof Downloads</strong>: Because issue descriptions are untrusted user input, we block risky URLs (like internal network IPs or cloud metadata endpoints), cap downloads at 20 MB, and fall back cleanly to text triage if an image is missing or unreachable.</li></ul><h4>D. Deterministic Image Downscaling (Saving Your Token Budget!)</h4><p>Here’s a classic “gotcha” from production: high-resolution screenshots and uncompressed PNGs can easily top several megabytes each. Shoving dozens of multi-megabyte images into a review prompt is a recipe for pain. Review request latency can spike towards GitHub Action timeout thresholds, you risk slamming into API request limits, and you waste tokens on pixels that add no extra value to the review.</p><p>To fix this, we added automated, deterministic downscaling using <strong>Pillow</strong>. Any image (.png, .jpg, .webp, .gif) larger than <strong>600 KB</strong> is compressed to a much smaller file size. Cool, right?</p><h4>2. Upgraded Default Model (gemini-3.8-flash) &amp; Thinking Budget Controls</h4><p>When I wrote Part 1, our default model was gemini-3.6-flash. Since then, we upgraded through gemini-3.7-flash and now default to <strong>gemini-3.8-flash</strong>.</p><p>We also introduced granular <strong>Reasoning &amp; Thinking Budget Control</strong> via the thinking_level input:</p><pre>- name: Run Gemini Review Action<br>        uses: derailed-dash/gemini-review-action@v1<br>        with:<br>          gemini_api_key: ${{ secrets.GEMINI_API_KEY }}<br>          github_token: ${{ secrets.GITHUB_TOKEN }}<br>          thinking_level: &#39;low&#39;   # Options: low, medium, high, off, or integer budget (e.g. 2048)</pre><p><strong>How thinking budgets work:</strong></p><ul><li><strong>Dynamic by Default (Recommended for PRs)</strong>: When omitted, Gemini runs in dynamic mode, automatically allocating reasoning tokens based on the complexity of your diff.</li><li><strong>FinOps Token Budget Limits (e.g. </strong><strong>low or </strong><strong>medium)</strong>: Because thinking tokens are billed at the higher <strong>output token rate</strong>, cost-conscious teams can cap the maximum reasoning headroom to ensure predictable per-review spending on massive diffs.</li><li><strong>Blazing Fast Issue Triage (</strong><strong>thinking_level: &#39;off&#39;)</strong>: When running in triage mode (command: &#39;triage&#39;), the model only needs to classify an issue and pick labels. Deep reasoning is complete overkill here. Setting thinking_level: &#39;off&#39; forces thinking_budget=0, slashing triage latency to near-instant and cutting triage costs to fractions of a cent!</li></ul><h4>3. Smarter Sparse Mode: Bounded Dynamic Context Selection</h4><p>In Part 1, when a repository exceeded 1.5 MB (<em>Sparse Context Mode</em>), the action attached the directory tree and static core files (README.md, pyproject.toml, etc.), and left it at that.</p><p>But what if your PR modifies, say, src/payments/stripe_webhook.py inside a 50 MB monorepo, but you also have relevant additional files in the preo, like src/payments/models.py and tests/payments/test_stripe.py?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HrO6KzdqRL5-vbKl9aTBUQ.png" /><figcaption>Monorepo Context &amp; Self-Healing Engine</figcaption></figure><p>To solve this, we introduced <strong>Bounded Dynamic Context Selection</strong> (gemini_review/context.py and gemini_review/repo.py):</p><ol><li><strong>Prioritised Rule Files &amp; Core Context Budget Cap (</strong><strong>max_core_context_bytes)</strong>: In Sparse Mode, the action loads your core project documents. But what if a repository has a 2 MB README.md or dozens of markdown files? To prevent bulky documentation from blowing your token budget before code is even evaluated, we cap total core context at <strong>500 KB</strong> and enforce strict intake ordering:<br>- <strong>Agent rules take precedence</strong>: Coding instructions and team guidelines (AGENTS.md, GEMINI.md, CLAUDE.md, .cursorrules) are prioritised at the very front of the queue so the reviewer always respects your project&#39;s rules.<br>- <strong>Manifests &amp; architecture next</strong>: Core build manifests (pyproject.toml, package.json, go.mod) and root docs (README.md, ARCHITECTURE.md) fill the remaining budget up to the 500 KB ceiling.</li><li><strong>Two-Stage Dynamic Sister Context Selection</strong>: In a 10,000-file monorepo, feeding every file path to an LLM wastes massive tokens before the review even starts. So we built a two-stage funnel:<br>- <strong>Stage 1 (Deterministic Ranking)</strong>: A Python ranker parses the PR diff for import statements, directory proximity, and sister test/implementation filenames to bound the candidate pool to the top 500 files (max_candidate_files).<br>- <strong>Stage 2 (4-Tier Architectural Selection)</strong>: A fast, structured call to Gemini picks up to <strong>8 of the highest-value sister files</strong> from that pool across four architectural tiers: <em>Tier 1</em> (direct import dependencies and base classes), <em>Tier 2</em> (sister unit tests, fixtures, and schemas), <em>Tier 3</em> (domain precedents in sibling modules), and <em>Tier 4</em> (shared utilities and configuration helpers).</li><li><strong>Static Override (</strong><strong>extra_context_files)</strong>: If you already know specific files you <em>always</em> want attached alongside every review, you can pass extra_context_files: &#39;src/core/security.py,docs/api-spec.pdf&#39; and bypass the dynamic selector call entirely!</li></ol><h4>4. Extending Guardrails vs Replacing Prompts + Curated Skills Synchroniser</h4><p>Originally, if you wanted to add a single team rule — like <em>“All SQL queries must use parameterised inputs”</em> — you had to copy the entire .github/commands/gemini-review.toml file into your repo. The problem? As soon as you forked the base prompt into your repo, you missed out on all the upstream prompt engineering improvements we shipped in later versions (like our <strong>5-Axis Quality Evaluation Framework</strong>)!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CGpB9clk10dnfTkN7_u37g.png" /><figcaption>AI Code Reviewer: Upgraded Intelligence Architecture with Skills &amp; MCP</figcaption></figure><p>The answer is to cleanly separate <strong>extending</strong> instructions from <strong>replacing</strong> them:</p><ul><li><strong>Extend (Recommended)</strong>: Simply drop a Markdown file at .github/review-instruction-additions.md in your repo, or pass inline rules via the custom_instructions workflow input:</li></ul><pre> with:<br>    custom_instructions: |<br>      - Ensure all database queries use parameterised inputs.<br>      - Verify that new API endpoints include Pydantic request validation.</pre><ul><li>The action keeps the full built-in 5-axis quality rubric, persona overlays, and schema rules, and cleanly appends your custom guardrails!</li><li><strong>Equip Custom Agent Skills (</strong><strong>.agents/skills/)</strong>: Beyond simple markdown rules, you can mount full Agent Skills directly into your repository. The action automatically discovers skills placed in .agents/skills/ — whether they are custom coding guidelines, security checks, or official skills from the google/skills catalog — and equips Gemini to enforce those exact domain practices during reviews.</li><li><strong>Bundled One-Command Installer Skill</strong>: I had already created an agent skill called install-gemini-code-review-action, in my <a href="https://github.com/derailed-dash/dazbo-agent-skills">dazbo-agent-skills</a> repo. Now, we also mirror this skill directly inside our reviewer repo. You can install the skill straight from the action repo:</li></ul><pre>npx skills add derailed-dash/gemini-review-action -y -g --skill install-gemini-code-review-action</pre><p>Once installed, simply prompt your agent: <strong>“Set up Gemini code review in this repository.”</strong> The agent detects any existing workflows, interviews you on your preferred model, auth method (WIF vs API key), and persona, and generates the workflow files for you!</p><h4>5. Workflow Polish: Auto-Resolving Threads &amp; Filtering Competing Bots</h4><p>Two of my favourite quality-of-life improvements came directly from real-world friction on busy pull requests:</p><h4>A. Auto-Resolving Addressed Review Threads</h4><p>When you push a fix to a PR, the action re-examines prior review comments, confirms the issue has been resolved, and marks it fixed.</p><p>However, if your repository enforces GitHub’s <strong>“Require conversation resolution before merging”</strong> rule, someone still has to manually click “Resolve conversation” on every single thread before you can merge — even though the bot already agreed you fixed it!</p><p>With resolve_addressed_threads: &#39;true&#39;, the action automatically clicks &quot;Resolve&quot; on its own threads as soon as Gemini verifies the fix. It stays strictly in its lane: it only touches threads the bot created, never touches human comments, and gracefully warns you if your GitHub token needs additional permissions to resolve conversations.</p><h4>B. Excluding Another Bot’s Comments (exclude_comment_authors)</h4><p>What happens if your team runs <strong>two</strong> AI reviewers on the same repo (for instance, claude[bot] alongside gemini-review-action)?</p><p>Because our action inspects existing PR comments to avoid repeating itself, it was picking up the other bot’s review notes too. That meant Gemini ended up reading the other bot’s homework and parroting the exact same points! You lost your independent second opinion, and you paid for hundreds of extra prompt tokens just to re-digest the other bot’s prose.</p><p>Adding exclude_comment_authors: &#39;claude[bot]&#39; filters out specified bots before building the prompt. Human feedback stays intact, and Gemini delivers a genuine, unbiased second opinion.</p><h3>Battle-Hardened in Production: War Stories &amp; Bug Fixes</h3><p>Here are a few real-world production bugs we hit since Part 1 — and how we engineered permanent fixes for them.</p><h4>War Story 1: The Generated Schema That Blew the Token Budget</h4><ul><li><strong>The Problem</strong>: A pull request touching 30 files crashed without posting a review, failing with a token limit error. The culprit? A tiny 4 KB diff buried inside a 2.9 MB auto-generated openapi.json file. Because our prompt builder attached the entire current content of every modified file alongside its diff, that single generated file consumed the entire Gemini token window on its own!</li><li><strong>The Fix</strong>: Rather than playing whack-a-mole adding filenames to an ignore list, we capped all attached source files at <strong>128 KB</strong> with a clean truncation marker (diff patches themselves are never capped). If a massive prompt still approaches the model’s budget, the action gracefully sheds peripheral repo context first, ensuring the core diff review always succeeds.</li></ul><h4>War Story 2: GitHub’s Concurrency Trap (and How Our Bot Caught It)</h4><ul><li><strong>The Problem</strong>: If a developer pushes three commits in quick succession, you don’t want to pay Gemini to review all three. We added a standard GitHub Actions concurrency block with cancel-in-progress: true to kill redundant runs. But we hit a snag: whenever someone left an ordinary comment on the PR, GitHub cancelled the running review! Workflow-level concurrency evaluates before the job filters kick in, treating any comment as a new run.</li><li><strong>Caught by Our Own Bot</strong>: We tried tweaking the concurrency key, but when we opened the pull request, <strong>our own Gemini Review Action reviewed it and pointed out our fix created a new race condition in the opposite direction!</strong></li><li><strong>The Fix</strong>: Moving concurrency down to the <strong>job level</strong> ensures GitHub only cancels runs for actual review triggers on the same PR:</li></ul><pre>jobs:<br>  review:<br>    concurrency:<br>      group: gemini-review-${{ github.event.pull_request.number || github.event.issue.number }}<br>      cancel-in-progress: true</pre><h4>War Story 3: Making “Commit Suggestion” Actually Work</h4><ul><li><strong>The Problem</strong>: Developers love clicking GitHub’s <strong>“Commit suggestion”</strong> button directly from an inline review comment. But LLMs are notorious for trimming leading whitespace or getting multiline replacement ranges slightly wrong — meaning a one-click commit can easily botch your indentation or overwrite the wrong lines.</li><li><strong>The Fix</strong>: Before posting an inline suggestion, we run a deterministic sanitiser over it. It matches the source file lines, restores stripped indentation, and corrects any off-by-one replacement ranges. When a developer clicks that button, the resulting commit is clean, properly indented, and builds first time.</li></ul><h4>War Story 4: Surviving Transient GitHub Outages</h4><ul><li><strong>The Problem</strong>: During an incident where the GitHub API threw intermittent 503 Service Unavailable errors, the action successfully generated reviews from Gemini, only to crash when posting them to GitHub. We’d paid for the tokens, but the reviews were lost into the ether.</li><li><strong>The Fix</strong>: We wrapped all GitHub comment and review posting in an exponential backoff retry. A review you’ve already paid for is never thrown away over a temporary API hiccup.</li></ul><h3>Wrapping Up</h3><p>Looking back at how far <a href="https://github.com/derailed-dash/gemini-review-action"><strong>Gemini PR Review &amp; Triage Action</strong></a> has come since that first weekend build in July, I’m genuinely blown away. What started as an emergency lifeboat after the sunsetting of run-gemini-cli has grown into a comprehensive, multimodal, enterprise-grade engineering harness.</p><p>Whether you’re a solo hobbiest who wants Emperor Palpatine or Rick Sanchez roasting your Python diffs for $0.02 a pop, or an Enterprise Cloud Architect rolling out keyless WIF authentication, BigQuery billing attribution, and architectural guardrails across hundreds of corporate repositories — this action is built for you:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*G2m6BP8NC22l7AE1y53eVw.png" /><figcaption>Rick Sanchez Persona Review Output</figcaption></figure><h3>Beyond CI/CD: Bringing Gemini Review into Your Agentic Loops</h3><p>While we designed this action primarily to automate GitHub pull request reviews, the modular architecture we’ve engineered under gemini_review/ unlocks an exciting new frontier: <strong>inner-loop agent verification</strong>.</p><p>In my recent deep dive, <a href="https://dev.to/gde/loop-engineering-in-the-harness-autonomous-self-correcting-systems-part-2-155e"><strong>Loop Engineering in the Harness: Autonomous, Self-Correcting Systems</strong></a>, I explored how reliable agentic architectures rely on tight, deterministic feedback loops (Generate → Verify → Correct → Converge) to eliminate hallucinated code and regressions before human intervention.</p><p>Because gemini_review is structured as a standalone, lightweight Python package executable via uv, you don&#39;t have to wait for a remote CI push to run it. You can wire the review engine directly into your local development harness as a pre-commit or pre-push verification step. Your local coding agent can execute a local review pass against its own unstaged git diff, parse the structured severity findings, and self-correct regressions with zero session bias — before anyone else ever inspects the code!</p><h3>Install It!</h3><p>If you haven’t tried it yet (or if you installed v1.1 back in July and haven&#39;t looked at the new inputs since!), now is the perfect time to upgrade:</p><ol><li><strong>Equip your AI agent in seconds</strong>, then prompt it: <em>“Set up Gemini code review in this repo”</em>:<br>npx skills add derailed-dash/gemini-review-action -y -g --skill install-gemini-review-action</li><li><strong>Or grab the starter workflows</strong> directly from <a href="https://github.com/derailed-dash/gemini-review-action">derailed-dash/gemini-review-action</a>.</li><li><strong>Give the repository a star ⭐ on GitHub</strong> — it massively helps others discover the project, and it helps me too!</li></ol><p>Let me know in the comments how you get on, what persona you’re running in your pipelines, and what features you’d love to see next!</p><h3>Before You Go</h3><ul><li><strong>Please share</strong> 📢 this with anyone that you think will be interested. It might help them, and it really helps me!</li><li>Please <strong>give me loads of claps</strong>! 👏 (Just hold down the clap button.)</li><li>Please <strong>leave a comment</strong> 💬. Interaction is good!</li><li><strong>Add a star</strong> ⭐ on my repos!</li><li><strong>Follow</strong> 👉 and <strong>subscribe</strong> 🔔, so you don’t miss my content.</li></ul><h3>References and Useful Links</h3><ul><li><a href="https://dev.to/gde/automated-github-code-reviews-using-google-gemini-4l0f">Part 1: Automated GitHub Code Reviews Using Google Gemini (Dev.to)</a></li><li><a href="https://medium.com/google-cloud/automated-github-code-reviewsusing-google-gemini-7b4d027b3092">Part 1: Automated GitHub Code Reviews Using Google Gemini (Medium)</a></li><li><a href="https://github.com/derailed-dash/gemini-review-action">Dazbo’s Gemini PR Review &amp; Triage Action (GitHub Repository)</a></li><li><a href="https://github.com/marketplace/actions/dazbo-s-gemini-code-review-issue-triage">Dazbo’s Gemini PR Review &amp; Triage Action (GitHub Marketplace)</a></li><li><a href="https://github.com/derailed-dash/dazbo-agent-skills">Dazbo’s Agent Skills Repository (</a><a href="https://github.com/derailed-dash/dazbo-agent-skills">dazbo-agent-skills)</a></li><li><a href="https://cloud.google.com/iam/docs/workload-identity-federation-with-deployment-pipelines?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Google Cloud Workload Identity Federation with GitHub Actions</a></li><li><a href="https://cloud.google.com/vertex-ai/generative-ai/pricing?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Google Cloud Vertex AI Generative AI Pricing</a></li><li><a href="https://antigravity.google/?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Google Antigravity</a></li><li><a href="https://dev.to/gde/from-vibing-chaos-to-reliable-agents-the-what-why-and-how-of-harness-engineering-4ba0">Beyond Vibe Coding: From Vibing Chaos to Reliable Agents — Harness Engineering Part 1</a></li><li><a href="https://dev.to/gde/loop-engineering-in-the-harness-autonomous-self-correcting-systems-part-2-155e">Beyond Vibe Coding: Loop Engineering in the Harness — Autonomous, Self-Correcting Systems (Part 2)</a></li><li><a href="https://dazbo.co.uk/">Dazbo’s Portfolio</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e3459f621aea" width="1" height="1" alt=""><hr><p><a href="https://medium.com/google-cloud/automated-github-code-reviews-using-google-gemini-update-enterprise-ready-multimodal-and-e3459f621aea">Automated GitHub Code Reviews Using Google Gemini Update: Enterprise-Ready, Multimodal, and…</a> was originally published in <a href="https://medium.com/google-cloud">Google Cloud - Community</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Loop Engineering in the Harness: Autonomous, Self-Correcting Systems]]></title>
            <link>https://medium.com/google-cloud/loop-engineering-in-the-harness-autonomous-self-correcting-systems-95d3ac8f64cf?source=rss-58a0bd8ba726------2</link>
            <guid isPermaLink="false">https://medium.com/p/95d3ac8f64cf</guid>
            <category><![CDATA[harness-engineering]]></category>
            <category><![CDATA[google-gemini]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[loop-engineering]]></category>
            <dc:creator><![CDATA[Dazbo (Darren Lester)]]></dc:creator>
            <pubDate>Thu, 08 Oct 2026 03:20:00 GMT</pubDate>
            <atom:updated>2026-10-24T17:13:05.251Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Part 2 of the series “Beyond Vibe Coding: The Engineering Blueprint for Reliable AI Agents”</em></p><blockquote>🗺️<strong> The Series Roadmap: Beyond Vibe Coding</strong></blockquote><blockquote>In 2026, software development is drowning in fragmented buzzwords: Harness Engineering, Loop Engineering, Context Engineering, Spec-Driven Development, Evaluators, and Sandboxes. Are they competing frameworks? Redundant prompt techniques? Or something fundamentally new?</blockquote><blockquote>This three-part series unifies these concepts into a single, cohesive architectural blueprint for building software with AI without writing code manually:</blockquote><ul><li><strong><em>Part 1: </em></strong><a href="https://dev.to/gde/from-vibing-chaos-to-reliable-agents-the-what-why-and-how-of-harness-engineering-4ba0"><strong><em>From Vibing Chaos to Reliable Agents: The What, Why, and How of Harness Engineering</em></strong></a><em><br>We diagnose the hangover of vibe coding, unpack the paradigm shift of shifting developer interventions left, establish our bedrock formula (</em><strong><em>Agent = Model + Harness</em></strong><em>), explore the Lewis Hamilton motorsport analogy, and deconstruct the deterministic components that make up a production developer environment.</em></li><li><strong><em>Part 2: Loop Engineering in the Harness: Autonomous, Self-Correcting Systems (This Article)</em></strong><em><br>We fire up the dynamic runtime engine of the harness: deconstructing the core loop components, establishing hybrid evaluation rubrics (deterministic tests + LLM goal fidelity + tool trajectory verification), and applying the Ratchet Principle.</em></li><li><strong><em>Part 3: Spec-Driven Development For-the-Win: Disposable Code, Living Blueprints, and Building Like a Boss (Coming Soon)</em></strong><em><br>We explore Spec-Driven Development (where code is disposable and specifications are the single source of truth), master Google Antigravity’s unified harness and interactive slash commands (</em><em>/grill-me, </em><em>/plan, </em><em>/goal, </em><em>/boost, </em><em>/learn), and combine formal specs with tactical workflows via the 4-Phase Hybrid Power Pattern.</em></li><li><strong><em>Part 4: Scaling the Harness: Practical Loop Patterns, ADK 2 Graphs, Recursive Harness², and Harness Maintenance (Coming Soon)</em></strong><em><br>We bring everything together with practical loop patterns you can wire into any repo (from declarative Markdown and </em><em>loopy to Google ADK 2 graphs), keep context clean across iterations with short-lived subagents, explore Google Research’s recursive Harness</em><strong><em>²</em></strong><em> framework, and cover ongoing harness maintenance.</em></li></ul><h3>The “Generate and Fix” Trap: Stop Being the Human Correction Loop</h3><p>Howdy friends, and welcome back!</p><p>In <a href="https://dev.to/gde/from-vibing-chaos-to-reliable-agents-the-what-why-and-how-of-harness-engineering-4ba0">Part 1</a>, we established the bedrock formula of modern agentic engineering:</p><p><strong>Agent = Model + Harness</strong></p><p>We deconstructed the core deterministic components of the harness: hierarchical context rules (AGENTS.md / GEMINI.md), living version-controlled state, curated Model Context Protocol (MCP) tools, on-demand skills, sandboxing, and observability. This is how we engineered the race car chassis.</p><p><strong><em>But wait… How does the work actually get done?</em></strong></p><p>In simple, naive AI interactions, developers issue a prompt, get some output, and then typically course-correct. Maybe the developer copies the output from the terminal into the chat and asks for a fix.</p><p>Notice what is happening here: we have an <strong>iterative cycle of action, verification, and course-correction</strong>. You cannot escape it. But in an ad-hoc vibe coding workflow, <strong>you are the loop</strong>. You act as an exhausted organic clipboard, forcing yourself to micromanage, inspect, and babysit every single turn. If you’re doing that, you haven’t automated much; you’re just paying a subscription to watch a terminal like a nervous ferret.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uSGq5Qw4wAeaN65vFMoY0g.png" /><figcaption>Watching the agent like a nervous ferret</figcaption></figure><p><strong>Loop Engineering</strong> automates that execution cycle.</p><p>Instead of forcing a human to watch and correct every turn, the loop enables our agent to cycle through action, verification, and course-correction <strong>autonomously within the harness</strong>. The agent executes the code, runs the automated test suites, evaluates its work against strict rubrics, refines its hypotheses, and makes corrections… On its own!</p><p>Crucially, this does not mean ejecting the human entirely. It means moving from continuous babysitting to strategic governance: <strong>Human-in-the-Loop does NOT mean Human-in-Every-Turn</strong>. The agent executes the rapid, deterministic feedback cycle autonomously, and only pauses to escalate to the human conductor at <strong>predefined escalation points</strong> — such as hitting persistent evaluation plateaus, encountering ambiguous requirements, or proposing irreversible, high-risk actions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*y395cFDH0JwTJpqC8NVMxA.png" /><figcaption>The Loop</figcaption></figure><blockquote><strong>A Crucial Distinction: The Development Workbench, Not Runtime Infrastructure</strong></blockquote><blockquote>When we talk about loop engineering, the goal is not usually to ship autonomous, self-modifying while-loops into a live, customer-facing production application. The value of harness engineering and loop engineering is that it provides a disciplined feedback mechanism during the development phase to <strong>build software that is production-grade</strong>. The loop is the automated forge where code is hammered, tested, and tempered; what you actually deploy to the cloud is resilient, production-ready software.</blockquote><blockquote>In tokenomics terms, this aligns to so-called <strong>Spec Spend</strong> (as explored in my upcoming enterprise whitepaper, The Agentic Revolution). Burning tokens in the developer harness to explore hypotheses, run tests, and self-correct is a <strong>high-value, high-leverage investment</strong>. Spending a couple of dollars in developer Spec Spend to forge a hardened application prevents catastrophic regressions and uncontrollable costs in production <strong>Run Spend</strong>.</blockquote><h3>Core Components of an Engineered Developer Loop</h3><p>To move beyond fragile trial-and-error, an engineered agentic loop requires distinct, interconnected components:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LqsU38ZEb6woC7imuCffQw.png" /><figcaption>Core Components of the Loop</figcaption></figure><p><strong>1 — Goal (The Anchor):</strong> A clear, unambiguous statement of what the agent must accomplish. In long-running autonomous runs, context compression can cause the model to drift. <strong>A resilient harness restates the core goal on every single iteration</strong>, keeping the agent firmly anchored to its original objective.</p><p><strong>2 — Context (The Environment &amp; Grounding Data):</strong> Everything the agent knows about the world before taking an action. This context extends far beyond prompt templates; it is the total information environment available to the loop:</p><ul><li><strong>Standing Architectural Rules &amp; Guardrails:</strong> Hierarchical instructions, coding standards, and operational boundaries drawn from GEMINI.md, AGENTS.md, and activated Agent Skills.</li><li><strong>Workspace Data &amp; Local Fixtures:</strong> Source code, schema files, seed datasets, sample payloads, and test mocks residing directly in the repository.</li><li><strong>Dynamic Knowledge &amp; Database State:</strong> Grounding information retrieved on demand via MCP tools or APIs, such as querying living documentation from (say) Confluence or Google Drive, inspecting issues on GitHub, or introspecting schemas and retrieving data from databases.</li></ul><blockquote>💡 <strong>Decoupling Engineering Guardrails from Application Specs</strong></blockquote><blockquote>A crucial architectural discipline in production harnesses is separating the <strong>Application Goal Specification</strong> from the <strong>Engineering Discipline</strong>. Naive prompt engineering often conflates the two, burying feature requirements beneath paragraphs of repetitive coding hygiene instructions. In reality, engineering guardrails (e.g. Python standards, mandating TDD approaches, Pydantic input parameterisation) should be <strong>independent of any single application spec</strong>. They represent reusable, global, organisation-, domain- or team-level engineering policies.</blockquote><blockquote>Centralising them in persistent workspace configurations (GEMINI.md, AGENTS.md) and reusable Agent Skills means you define them once and govern them centrally across all projects. The feature specification stays lean and domain-focused.</blockquote><p><strong>3 — Tools (The Levers):</strong> The restricted set of actions the agent is permitted to execute, including sandboxed shell commands and MCP server tools. The harness strictly limits tool access to reduce blast radius and limit tool choice paralysis.</p><p><strong>4 — Evals (The Objective Judge):</strong> The scoring mechanisms that evaluate the progress of each iteration. Without automated evals, the agent is flying blind. A robust harness evaluates across three dimensions:</p><ul><li><strong>Deterministic Output Checks:</strong> Mechanical unit tests, integration suites, and static linters (pytest, ruff, codespell) that verify the compilation, syntax, and functional correctness of the generated artifacts.</li><li><strong>Tool Trajectory &amp; Process Verification:</strong> Inspecting <em>how</em> the agent reached the result. Did it invoke only approved tools in the expected sequence? Did it respect tool schemas without hallucinating arguments? Did it avoid thrashing loops or calling forbidden APIs?</li><li><strong>Semantic LLM-as-a-Judge Analyses:</strong> Using models to evaluate the qualitative requirements — crucially including <strong>Goal Fidelity</strong> (verifying whether the solution actually delivers what was requested) and architectural cleanliness. While semantic autoraters can use rubric scales or qualitative feedback, framing them as <strong>binary yes/no assertions</strong> is often a very effective approach to eliminate subjective score drift.</li></ul><p><strong>5 — Memory (Loop History &amp; Scratchpad):</strong> The structured log of what has been attempted across preceding iterations. The agent records what worked, what failed, and the progression of evaluation scores over time. In a well-engineered harness, this persistent scratchpad prevents the model from repeating known dead-ends or thrashing in circular loops.</p><p><strong>6 — Iteration Discipline (The One-Change Rule):</strong> This is about applying the scientific method to autonomous iteration. This discipline applies specifically to changes evaluated for <strong>semantic value</strong>. When iterating against semantic autoraters, the loop instructs the agent to <strong>modify one thing at a time</strong> before re-scoring. If multiple semantic changes are introduced at once, causal observability collapses: if a rubric score drops, the agent cannot tell which alteration caused the regression. Conversely, purely mechanical deterministic fixes (resolving compiler syntax, fixing lint warnings from ruff or codespell, and formatting) can and should be batch-remediated in a single turn.</p><p><strong>7 — Escalation Protocol (The Safety Valve):</strong> Explicit operational triggers that halt the loop and request human intervention. For example, if an agent encounters conflicting specifications, hits three consecutive evaluation failures, or proposes an irreversible action (e.g. deleting cloud infrastructure), it must not hallucinate a guess. It must package its findings, present a reasoned recommendation, and yield control to the human conductor.</p><p><strong>8 — Stopping Condition (The Finish Line):</strong> Clear, programmatic exit criteria that tell the loop when execution is complete. Such as:</p><ul><li>100% of deterministic checks and semantic rubric assertions passing with zero lint warnings.</li><li>Target state verification: the critique model explicitly certifies that the initial goal has been fully met.</li><li>A convergence plateau where successive iterations produce negligible score improvements.</li><li>Hard circuit breakers, such as reaching a maximum iteration count (e.g. 5 loops) or a predefined token spend limit.</li></ul><p><strong><em>Is It a Loop Bug or a Harness Bug?</em></strong><em> When an autonomous agent goes off the rails during development, we often waste days trying to fix the wrong things.</em></p><p><em>Use this simple litmus test:</em></p><ul><li><strong>If you can resolve the issue by refining the prompt, sharpening the goal, updating the rubric, or adjusting the stopping condition → It is a Loop bug.</strong></li><li><strong>If resolving the issue requires changing tool definitions, adjusting sandbox permissions, modifying execution timeouts, or intercepting runtime errors → It is a Harness bug.</strong></li></ul><p><em>Don’t attempt to solve a harness problem by writing a more pleading prompt. Fix the layer that actually broke!</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9GBh0WJoa_5Fbh-xeCfKlQ.png" /><figcaption>Loop bug vs harness bug</figcaption></figure><h3>The Ratchet Principle: Engineering Continuous Improvement</h3><p>What happens once you diagnose an edge case or failure that <em>wasn’t</em> caught by your initial harness?</p><p>This brings us to the operational crown jewel of harness engineering: <strong>The Ratchet Principle</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yQAV_01e4owYtkZGTVEZ7A.png" /><figcaption>Ratchet Principle</figcaption></figure><p>When an AI agent stumbles, makes an architectural mistake, or introduces a subtle bug, the amateur response is to re-prompt the model in the chat: <em>“No, don’t do that, use the new API method instead.”</em></p><p>That is an utter waste of a lesson. The moment that conversation ends, that context evaporates into thin air. In tomorrow’s session, on a new branch, or when a teammate runs a similar task, the agent will walk directly into the exact same trap.</p><p>Under the Ratchet Principle:</p><blockquote><strong><em>Whenever an agent makes a mistake, you update the harness, not just the prompt.</em></strong></blockquote><ul><li><strong>Did the agent use an outdated API parameter?</strong> Update the workspace GEMINI.md or add an authoritative type schema.</li><li><strong>Did the agent forget to handle a specific error code?</strong> Add a permanent check to the binary evaluation rubric.</li><li><strong>Did the agent create an unformatted commit?</strong> Add a pre-commit hook to your Git harness.</li><li><strong>Did the agent struggle with a Google Cloud product?</strong> Load the corresponding skill from the google-cloud-developer plugin.</li></ul><p>Like a mechanical ratchet, every single failure permanently tightens the system. The harness clicks forward one notch. Once a rule or rubric check is committed to your repository, that specific failure mode is eliminated forever. Over time, your development harness becomes an impenetrable repository of institutional engineering wisdom.</p><h3>Why Loops Work: The Power of Feedback</h3><p>Why go through the effort of building a multi-component loop instead of relying on a single, well-crafted prompt?</p><p>I created an application that demonstrates building an application agentically, and compares <em>with and without</em> harness and loop, side-by-side.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WCycme3xzu8p38816bEFhA.png" /><figcaption>Side-by-side harness engineering demo</figcaption></figure><p>We’ll deep-dive into this shortly. But for now, I’ll give you the key findings from this application, which align with industry findings. Essentially, if you have a harness with a loop, these are the outcomes:</p><ol><li>You use <strong>fewer tokens.</strong></li><li>The <strong>overall time</strong> to get to a solution is <strong>dramatically reduced</strong>.</li><li>The <strong>human input</strong> required is <strong>massively reduced</strong>.</li><li>The <strong>overall cost</strong> to produce the solution is <strong>significantly reduced</strong>.</li><li>The <strong>quality</strong> of your solution, and its alignment to your requirements, is much better.</li><li>Most importantly — <strong>the solution is more likely to <em>actually work</em>.</strong></li></ol><p>TL;DR — single-pass prompting is gambling that an LLM will get dozens of interconnected variables right on the first try.</p><p>In contrast, wrapping the model in an automated loop mirrors the exact rhythm experienced software engineers use every day: propose a targeted change, run the tests and evals, feed the diagnostics straight back into the next turn, and iterate until the solution passes. The loop keeps the model honest, converges on working code dramatically faster, and stops intermediate errors from compounding into an unrecoverable mess.</p><h4>Three Traps to Avoid</h4><p>However, naive loops without rigorous harness constraints may fall prey to some common <strong>loop traps</strong>:</p><p><strong>1 — The Premature Exit:</strong></p><p>The agent runs superficial tests, assumes everything works because there were no syntax errors, and prematurely declares victory.<br><em>Neutralised by: Mandatory TDD gates and non-negotiable passing thresholds on automated test suites.</em></p><p><strong>2 — The Sycophantic Yes-Agent</strong></p><p>We all know that LLMs can be hopelessly sycophantic.<br><em>Q: “Do I look good in this shell suit?”</em><br><em>A: “Why yes, Dazbo! You look so hunky wearing that.”</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/768/1*vA2NDxdZI_0_llWghozpLw.png" /><figcaption>Do I look good in this shell suit?</figcaption></figure><p>When using an LLM to rate the output of a loop, the model easily falls into the same trap — uncritically praising the worker agent’s output.<br><em>Neutralised by: Structuring evaluations as strict, binary assertions and verifiable code artifacts.</em></p><p><strong>3 — The Bottomless Cash Pit</strong></p><p>An unachievable rubric threshold or a brittle stopping condition traps the agent in a recursive loop, repeatedly modifying the same file and burning through hundred-dollar token budgets without making forward progress. While developer Spec Spend is meant to be exploratory, an unconstrained loop turns healthy investment into an uncontrolled financial bonfire.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IHTKoBLcHXl7OGuzjzVnOA.png" /><figcaption>Uncontrolled token burn</figcaption></figure><p><em>Neutralised by: Stall detection, plateau convergence checks, and hard iteration circuit breakers.</em></p><h3>The Hybrid Evaluation Rubric (Deterministic + Semantic)</h3><p>One of the biggest mistakes teams make when building agent loops is relying on fluffy, subjective evaluation prompts: <em>“Rate the quality of this code on a scale of 1 to 5.”</em></p><p>Models are notoriously inconsistent when asked for qualitative ratings; what scores a 4 on Tuesday might score a 2 on Wednesday. In contrast, <strong>binary rubrics</strong> convert these fluffy judgments into repeatable deterministic results. So let’s just turn the evaluations into verifiable yes/no questions!</p><p>Furthermore, as Google AI leaders like Shubham Saboo have pointed out in <a href="https://www.youtube.com/watch?v=ew6gBJNzC5w"><em>Finally, Loop Engineering Explained</em></a> when contrasting loops with traditional prompt engineering, reliable loops must <strong>automate what is checkable</strong> — and that begins with <strong>testing whether the generated output actually achieves the goal that was set</strong>.</p><p>The goal is not just passing unit tests and being clean of linting errors. We also need to be sure that the solution actually delivers the feature requested.</p><p>For this reason, our evaluation rubric should combine <strong>deterministic checks</strong> with <strong>semantic checks</strong>. The semantic checks are executed by a critique agent or LLM-as-a-judge autorater.</p><p>And we must evaluate <strong>tool trajectory</strong> alongside code execution. You cannot evaluate only the final artifact; the harness must verify <em>how</em> the agent reached the result. Tool trajectory evaluation operates across two complementary layers:</p><ol><li><strong>Deterministic Trajectory Evaluation (Mechanical Trace Inspection):</strong> Code-level validation inspecting the execution trace without using an LLM. Did the agent invoke only authorised tools from its allowlist? Did arguments strictly satisfy JSON parameter schemas? Did it follow mandatory prerequisite ordering (e.g. running the test suite <em>before</em> attempting to commit code)? Did it stay within maximum tool call budgets rather than entering an unconstrained loop?</li><li><strong>Semantic Trajectory Evaluation (Reasoning &amp; Efficiency):</strong> Evaluated by a critique agent or LLM judge. Did the agent select the <em>most appropriate and efficient</em> tool for the problem, or did it take an absurdly convoluted detour (e.g. running fifteen individual file-read calls instead of a single targeted grep)? When a tool returned an unexpected error or empty result, did the agent reason soundly about the failure and adapt its strategy, or did it flail repetitively?</li></ol><p>An agent that arrives at a working file after making ten redundant, unapproved tool calls or taking dangerous exploratory detours is an operational failure.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lVEU4BlT5Rdr6j4AW7gBQQ.png" /><figcaption>Two Tier Evaluation Rubrics</figcaption></figure><h3>The Anatomy of a Production-Grade Evaluation Rubric</h3><p>Here is a simple example; a 10-check rubric illustrating how deterministic and semantic checks — across both code and tool trajectories — work together:</p><pre># Evaluation Rubric<br><br>## Part A: Deterministic &amp; Static Checks (Automated via Test Runners, Linters &amp; Trace Inspectors)<br>| ID | Check (1 = True, 0 = False) | Evaluation Method | Result |<br>|:---|:---|:---|:---:|<br>| 1.1 | All unit tests in `tests` pass cleanly? | `pytest` runner | [ ] |<br>| 1.2 | `uvx ruff check .` and `uvx codespell` report zero errors or warnings? | Static analysis CLI | [ ] |<br>| 1.3 | Function signatures use PEP 585 standard type hints (e.g. `list[str]`, `dict[str, Any]`)? | Linter | [ ] |<br>| 1.4 | All database queries use parameterised inputs (no raw f-strings or string concatenation)? | Static analysis / AST | [ ] |<br>| 1.5 | **Deterministic Trajectory Conformance:** Did the agent invoke only authorised tools, satisfy parameter schemas, respect tool call limits, and follow mandatory prerequisite ordering (e.g. test before commit)? | Harness Trace Inspector (Code) | [ ] |<br><br>## Part B: Semantic Yes/No Analyses (Evaluated via LLM-as-a-Judge / Critique Agent)<br>| ID | Binary Semantic Assertion (1 = True, 0 = False) | Evaluation Method | Result |<br>|:---|:---|:---|:---:|<br>| 2.1 | **Target State Completion (Goal Fidelity):** Does the generated solution demonstrably deliver all functional capabilities and explicit deliverables declared in the goal statement, without omitting mandatory acceptance criteria? | Semantic Judge | [ ] |<br>| 2.2 | **Scope Discipline (Anti-Hallucination):** Does the solution achieve the goal without generating unrequested speculative abstractions, hallucinating unused helper modules, or modifying files outside the agreed boundary? | Semantic Judge | [ ] |<br>| 2.3 | **Trajectory Efficiency &amp; Tool Reasoning:** Did the agent choose sensible, direct tool pathways to solve the problem rather than taking convoluted detours or repeatedly retrying failed tool strategies without adaptation? | Semantic Judge | [ ] |<br>| 2.4 | **Architectural Integrity &amp; Interface Boundaries:** Does the code maintain clean boundaries, preventing internal database implementation details or raw connection strings from leaking into public interfaces or caller errors? | Semantic Judge | [ ] |<br>| 2.5 | **Design Rationale &amp; Intent:** Does the module-level docstring clearly explain the architectural *why* (why this pattern was chosen over alternatives) rather than merely restating what the code does? | Semantic Judge | [ ] |<br><br>**Passing Criteria:** 10 / 10 (100% pass required before code can be committed).<br>**Iteration Policy:** Failures on code, test suites, or sub-optimal tool trajectories feed diagnostic feedback into the next turn for autonomous remediation.<br>**Hard Stop (Human Escalation):** The harness only aborts and alerts the developer if the agent breaches security boundaries (e.g. unapproved destructive commands), enters an unrecoverable tool thrashing loop, or reaches the maximum iteration limit (5 loops).</pre><p>Notice how this structure enforces discipline across every dimension:</p><ul><li><strong>Checks 1.1–1.5</strong> verify mechanical hygiene (syntax, compilation, test passage, and deterministic tool trajectory allowlists/orderings).</li><li><strong>Checks 2.1–2.2</strong> verify <strong>Goal Fidelity</strong> — comparing the candidate output directly against the initial user intent. If the agent forgot batch processing or wandered off into speculative refactoring, Check 2.1 or 2.2 drops to 0, and the loop refuses to exit.</li><li><strong>Check 2.3</strong> verifies semantic tool trajectory efficiency — ensuring the agent didn’t burn tokens on irrational intermediate actions.</li><li><strong>Checks 2.4–2.5</strong> verify semantic architecture, error safety, and living documentation quality.</li></ul><p>By presenting the agent with this exact markdown rubric, the model can self-score its work after every turn. It doesn’t wonder whether its output is “good enough”; it knows with precision whether it satisfied the criteria and achieved the goal.</p><h4>No Arbitrary Point Ceiling</h4><p>While our example shows 10 checks for clear illustration, let’s avoid any misconceptions: <strong>an evaluation rubric does not have an arbitrary 10-point ceiling</strong>. In practice, your rubric can have whatever point total you want. For example, you might also include:</p><ul><li><strong>Frontend contract checks:</strong> Verifying TypeScript interface alignment, component contracts, and state store mutations.</li><li><strong>API wiring verification:</strong> Confirming route registration, middleware attachment, payload serialisation, and HTTP status code mappings.</li><li><strong>Latency bounds:</strong> Asserting that mock or integration calls stay strictly within defined latency budgets (e.g. 95% in under 250ms).</li><li><strong>Security invariants:</strong> Enforcing parameterised queries, sanitised inputs, absence of leaked credentials or stack traces, and zero-trust egress policies.</li></ul><h3>Side-by-Side Demo: The Harness Engineering Workbench</h3><p>Theory is cheap. Let’s prove it!</p><p>To demonstrate the quantifiable difference between raw vibe coding and disciplined loop engineering, I built an open-source, interactive, split-screen evaluation workbench: the <a href="https://github.com/derailed-dash/harness-engineering-demo">Harness Engineering Demo</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DaWsqSE2g1d0-JA1PZhXyQ.png" /><figcaption>Harness Engineering Workbench</figcaption></figure><p>The workbench pits two autonomous tracks against each other in real time, each running concurrently with the same model. (The demo uses Gemini 3.8-Flash out-of-the-box, but this is configurable.)</p><ol><li><strong>The Unharnessed Track — Vibe Coding:</strong> A raw single prompt containing the user specification is dispatched directly to the model. We supply no additional workspace rules or guardrails, and no specialised skills. There is no autonomous loop and no iterative self-healing. We evaluate the output once, at the very end.</li><li><strong>The Harnessed Track — Loop Engineering:</strong> An autonomous multi-turn loop executes in an isolated workspace. We provide the <em>same specification prompt</em> as before. But this time we also provide additional context and guardrails, we equip the agent with relevant skills, and we evaluate each iteration with a 14-point rubric.</li></ol><p><em>What is the demo actually building?</em> Well, the specification is for the non-trivial task of building a <em>“Cosmic Trivia &amp; Strategy Conquest Game”</em>, a full-stack sci-fi game featuring a FastAPI backend, a directed graph of galactic sectors, turn-based combat, dynamic AI trivia generation, and an interactive browser-based star map with real-time HUD metrics and combat dialogues.</p><h4>The Harness: Context, Skills, and Living Memory</h4><p>Let’s start by looking at the supporting <strong>context</strong> we supply. This would typically be part of your global reusable context (e.g. AGENTS.md or GEMINI.md) that you use as an engineer, as a team, or even as an organisation. For example:</p><pre># Engineering Context &amp; Harness Guardrails<br><br>## Persona &amp; Role<br><br>You are a Principal Software Engineer and Systems Architect. You produce modular, <br>production-grade, and self-documenting code following strict engineering standards. <br>You practice test-driven development, enforce strict interface contracts, <br>and design systems with clear separation of concerns.<br><br>## 1. Environment &amp; Coding Standards<br><br>- **Python Version**: Python 3.13 standard.<br>- **Type Hinting**: Strict PEP 585 (use built-in `list[str]`, `dict[str, Any]`, `tuple[int, ...]`, <br>  never use deprecated `typing.List` or `typing.Dict`).<br>- **Scope Discipline**: Only create required files specified in the task goal or architecture. <br>  Do not generate speculative helper files, hallucinated third-party dependencies, or dead code.<br><br>## 2. Test-Driven Development (TDD)<br><br>- **Mandatory Test Suite**: Author a comprehensive unit test suite in `tests/` before or <br>  alongside implementation covering core domain logic, edge cases, and state transitions.<br>- **Verification Gate**: Verify that all unit tests pass cleanly via `pytest` before <br>  declaring any task complete.<br><br>## 3. Backend Architecture &amp; Security Guardrails<br><br>- **Parameterisation**: All API request bodies and input parameters must use <br>  structured Pydantic models with input validation. Never accept raw unvalidated <br>  dictionaries or execute dynamic evaluation (`eval`, `exec`).<br>- **Error Handling**: Invalid inputs, illegal state transitions, or domain rule <br>  violations must trigger actionable HTTP 400 responses without leaking internal <br>  stack traces.<br>- **Architectural Integrity**: Maintain strict modular boundaries. Keep domain <br>  logic pure and independent of HTTP transports, database drivers, or UI presentation. <br>  Expose REST endpoints with explicit routing.<br>- **Documentation**: Include clear top-of-module docstrings explaining architectural <br>  intent and design rationale (&#39;why&#39; over &#39;what&#39;).<br><br><br>## 4. Autonomous Harness Workflow &amp; Tooling Protocol<br><br>- **On-Demand Skill Discovery &amp; Activation**: Inspect the `&lt;available_skills&gt;` <br>  catalogue in your harness context; when `# ACTIVATED ENGINEERING SKILLS` <br>  is already loaded in your context, apply those skills directly without <br>  re-activating them.<br>- **Initial Turn Synthesis (Iteration 1)**: Respond directly in Markdown with <br>  a concise **Architectural &amp; TDD Verification Plan** (150–250 words) detailing <br>  module boundaries, Pydantic request/response schemas, and the core unit test <br>  cases (`tests/`), concluding with a single brief (~15-line) Python code block <br>  illustrating the core Pydantic request/response contracts (omit full multi-file <br>  implementations).<br>- **Automated Gate Verification**: At the end of each iteration turn, your <br>  generated deliverables are materialised into the workspace and the harness <br>  evaluation gate executes `pytest`, `ruff`, and all mechanical and semantic <br>  rubric checks, recording a structured **Iteration Evidence Receipt** in <br>  `LIVING_MEMORY.md`.<br>- **Targeted Self-Healing from Living Memory (Iteration 2+)**: When living <br>  memory diagnostics from a prior iteration are present in your context, <br>  respond directly in Markdown with a concise **Self-Healing Remediation <br>  Summary** (100–200 words) citing each failed check ID, its root cause, and <br>  the targeted code fix applied.</pre><p>Note how this additional context defines the engineering persona, enforces our coding standards, mandates test-driven development (TDD), describes error handling requirements, captures how we want code documentation written, and establishes the autonomous loop workflow.</p><p>Also, because we want output of each loop to guide the next efficiently, this particular loop instructs the agent not to dump full multi-file implementations into the conversational stream. Instead it asks for a succinct summary of waht went wrong and how we might remediate next time.</p><p>Crucially, this is all <em>reusable, project-agnostic</em> loop context rather than context coupled to the application specification.</p><p>Next, observe that the harness has a number of <strong>specialised skills</strong> available to it:</p><ul><li>api-and-interface-design</li><li>gemini-api-dev</li><li>test-driven-development</li></ul><p>The agent will make use of these as required.</p><p>Finally, our harness has <strong>living memory</strong>: a persistent feedback register. When an iteration fails, the exact rubric diagnostics, failed check IDs, and remediation hints are recorded and reinjected into the next turn’s prompt, alongside the original specification. This is what enables surgical self-healing rather than blind guessing.</p><p>We do this by simply storing the output of each iteration in a markdown file called LIVING_MEMORY.md (or whatever you want to call it).</p><h4>Under the Hood: Orchestration the Loop with Google ADK</h4><p>To ensure a fair, apples-to-apples comparison, both tracks in the workbench are executed through the <strong>Google Agent Development Kit</strong>. This isolates the impact of the harness rather than comparing different SDK runtimes:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YLAkzVrYQ-uxasPe5MIKdg.png" /></figure><p><strong>Track 1 — Unharnessed (</strong><strong>app/agents/unharnessed.py):</strong> Runs a bare ADK Agent(instruction=&quot;&quot;, tools=[]) in a single pass. It receives only the raw specification from specs/cosmic_conquest_spec.md with no persona, no skills, and no verification tools.</p><p><strong>Track 2 — Harnessed (</strong><strong>app/agents/agent.py &amp; </strong><strong>app/agents/harnessed.py):</strong> Separates pure ADK agent definitions (agent.py, compatible with adk web, adk run, and adk eval) from the SSE runner (harnessed.py), orchestrating an ADK LoopAgent(max_iterations=5) composed of two sub-agents operating over shared session state (session.state):</p><ol><li><strong>Builder </strong><strong>LlmAgent (</strong><strong>harnessed_builder_agent):</strong> Equipped with our harness_context.md, the &lt;available_skills&gt; catalogue, five session-scoped ADK FunctionTools (activate_skill, write_workspace_file, read_workspace_file, run_pytest, run_ruff_check), and an after_agent_callback (materialise_builder_output_callback) that materialises generated files into the isolated workspace. Each tool operates on tool_context.state[&quot;workspace_path&quot;], injects activated skills into tool_context.state[&quot;skills_context&quot;], and appends deterministic audit entries to tool_context.state[&quot;tool_trajectory&quot;].</li><li><strong>Custom Gate Agent (</strong><strong>RubricEvaluationGateAgent(BaseAgent)):</strong> A pure evaluation and escalation gate that executes pytest and ruff, runs the 14-point rubric evaluator (loading semantic criteria declaratively from harness/RUBRIC.md), persists diagnostics to LIVING_MEMORY.md, and yields EventActions(escalate=should_escalate, state_delta=...). When escalate=True is emitted (on a 14/14 pass or plateau), the parent ADK LoopAgent terminates cleanly; otherwise, the updated living_memory_feedback state flows automatically into the builder agent&#39;s next iteration.</li></ol><h4>Programmatic Exit Conditions &amp; Circuit Breakers</h4><p>In autonomous loop engineering, you must never allow an agent to loop indefinitely or guess when it has finished. The workbench orchestrator enforces three strict programmatic exit conditions:</p><ol><li><strong>Target Gate Pass (Clean Exit)</strong>: To eliminate subjective “good enough” assessments, the workbench implements a rigorous, two-tier 14-point rubric schema. We will exit the loop upon achieving a flawless <strong>14 / 14 (100%)</strong> rubric score across all deterministic and semantic checks.</li><li><strong>Hard Iteration Cap (Spend Circuit Breaker)</strong>: A hard limit of <strong>5 iterations</strong> (MAX_ITERATIONS = 5). If the agent cannot satisfy all checks within five turns, execution halts immediately to prevent runaway token spend.</li><li><strong>Convergence Plateau (Stall Detection)</strong>: If the rubric score fails to improve across <strong>3 consecutive turns</strong> (PLATEAU_LIMIT = 3), the loop aborts early rather than thrashing in circular remediation attempts.</li></ol><p>Here is the exact production rubric configured in our workbench:</p><pre># Production Evaluation Rubric: Cosmic Conquest<br><br>## Tier 1: Deterministic Mechanical Sensors (8 Points)<br><br>| ID | Category | Check | Evaluation Method |<br>|:---|:---|:---|:---|<br>| 1.1 | Deterministic | Unit Tests Passage | `pytest` runner |<br>| 1.2 | Deterministic | Static Analysis (Ruff) | `ruff check .` |<br>| 1.3 | Deterministic | PEP 585 Type Hints | AST Inspector (`list[str]`, `dict[str, Any]`) |<br>| 1.4 | Deterministic | Parameterised Input Safety | AST Inspector (`eval()`, `exec()`) |<br>| 1.5 | Deterministic | Deterministic Tool Trajectory | Harness Trace Inspector |<br>| 1.6 | Deterministic | Interactive Frontend Client Structure | HTML &amp; DOM Inspector (`#shields`, `#energy`, combat modal) |<br>| 1.7 | Deterministic | Frontend Interactive API Wiring | AST / Script Inspector (fetch bindings to `/api/game/attack`, `answer`, `state`) |<br>| 1.8 | Deterministic | Dynamic API Contract &amp; Progression Topology | Dynamic TestClient Boot (FastAPI startup &amp; valid progression targets) |<br><br>## Tier 2: Semantic Goal Fidelity &amp; Ergonomics (6 Points)<br><br>| ID | Check | Evaluation Method | Weight |<br>|:---|:---|:---|:---:|<br>| 2.1 | Semantic | Target State Completion (Goal Fidelity) | Gemini LLM-as-a-Judge |<br>| 2.2 | Semantic | Scope Discipline (Anti-Hallucination) | Gemini LLM-as-a-Judge |<br>| 2.3 | Semantic | Architectural Integrity | Gemini LLM-as-a-Judge |<br>| 2.4 | Semantic | Error &amp; Remediation Clarity | Gemini LLM-as-a-Judge |<br>| 2.5 | Semantic | Design Rationale &amp; Intent | Gemini LLM-as-a-Judge |<br>| 2.6 | Semantic | Progression Ergonomics &amp; Interactivity | Explicit single-action sector card/tile affordance inspector |<br><br>Passing Bar: 14 / 14 (Non-negotiable 100% threshold)</pre><h4>The Head-to-Head Verdict</h4><p>If you run the side-by-side comparison multiple times, you’ll find the following:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fsiteproxy.ruqli.workers.dev%3A443%2Fhttps%2Fdatawrapper.dwcdn.net%2FBZ5Ux%2F5%2F&amp;display_name=Datawrapper&amp;url=https%3A%2F%2Fsiteproxy.ruqli.workers.dev%3A443%2Fhttps%2Fdatawrapper.dwcdn.net%2FBZ5Ux%2F5%2F&amp;image=https%3A%2F%2Fsiteproxy.ruqli.workers.dev%3A443%2Fhttps%2Fdatawrapper.dwcdn.net%2FBZ5Ux%2Fplain-s.png%3Fv%3D5&amp;type=text%2Fhtml&amp;schema=dwcdn" width="600" height="372" frameborder="0" scrolling="no"><a href="https://medium.com/media/0f88cb89b4996885871ebf34705d3f6d/href">https://medium.com/media/0f88cb89b4996885871ebf34705d3f6d/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BJyYb2VlOlAA8EDBueHJdQ.png" /><figcaption>Side-by-side results</figcaption></figure><p>You can watch a full live demonstration of the solution here:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fsiteproxy.ruqli.workers.dev%3A443%2Fhttps%2Fwww.youtube.com%2Fembed%2F7pKl2SvmPuU%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fsiteproxy.ruqli.workers.dev%3A443%2Fhttps%2Fwww.youtube.com%2Fwatch%3Fv%3D7pKl2SvmPuU&amp;image=https%3A%2F%2Fsiteproxy.ruqli.workers.dev%3A443%2Fhttps%2Fi.ytimg.com%2Fvi%2F7pKl2SvmPuU%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/454e53c726b07053e49ffc9dade2a0cb/href">https://medium.com/media/454e53c726b07053e49ffc9dade2a0cb/href</a></iframe><p><em>(You can inspect the full workbench codebase, explore the evaluation engine, or run the live demo locally. Just clone the open-source repository at </em><a href="https://github.com/derailed-dash/harness-engineering-demo"><em>derailed-dash/harness-engineering-demo</em></a><em>.)</em></p><h4>Unpacking the Results: Why the Harnessed Loop Dominates</h4><p>These numbers might appear counter-intuitive at first glance. Why would an iterative multi-turn loop be <strong>three to four times faster</strong> and <strong>around one-seventh of the cost (~85% cheaper)</strong> of a single-turn prompt?</p><p>Let’s break down these results:</p><h4>1. Only the Harnessed Version Actually Works</h4><p>The most sobering lesson of the comparison is functional: <strong>the vibe-coded solution is almost always non-functional</strong>.</p><p>On paper, the unharnessed agent produced a directory of plausible-looking files. It wrote Python classes and generated a colourful HTML file. But when loaded into a browser, the game was completely dead. The sector tiles lacked click event bindings, the combat modal was un-wired, and several FastAPI endpoints failed to resolve dynamic path parameters correctly.</p><p>It typically scores between 6 and 11 out of 14, failing key mechanical checks and progression requirements.</p><p>In contrast, the harnessed agent’s code <em>always </em>achieves a flawless 14 / 14. Because these checks include interrogating the frontend DOM structure and API bindings, we can be reasonably confident that the resulting game actually plays.</p><h4>2. Much Faster: The Paradox of Disciplined Bursts</h4><p>Intuition suggests that a single-turn prompt should be faster than an iterative loop. The workbench proves the exact opposite: <strong>the harnessed loop usually finishes in 24 to 35 seconds, whilst the unharnessed run invariably drags on for 1.5 to 2+ minutes.</strong></p><p>Why? Two architectural factors work together here:</p><ul><li><strong>Disciplined Output vs Speculative Rambling:</strong> When an unguided model is asked to build a full application in a single pass without constraints, it falls into speculative rambling. It attempts to stream 15,000–17,000+ output tokens in one continuous burst — generating repetitive boilerplate, guessing at auxiliary classes, and second-guessing its own structure mid-stream. The harnessed agent, by contrast, is guided by harness/harness_context.md and activated skills to operate in compact, high-signal bursts: Turn 1 synthesises a concise <strong>Architectural &amp; TDD Verification Plan</strong> (~800 output tokens, scoring 12.0 / 14.0 out of the gate) while materialise_builder_output_callback materialises the modular workspace files, and Turn 2 emits a focused <strong>Self-Healing Remediation Summary</strong> (~240 output tokens) to close the remaining +2.0 gap (12.0 -&gt; 14.0).</li><li><strong>Externalised Loop Reasoning (</strong><strong>ThinkingLevel.LOW):</strong> Rather than burning 30+ seconds per turn in unconstrained internal chain-of-thought pondering (thinking_budget is deprecated on Gemini 3.x+, so we configure ThinkingConfig(thinking_level=ThinkingLevel.LOW) on gemini-3.8-flash), the harnessed builder agent externalises its iterative reasoning into the deterministic ADK LoopAgent, the RubricEvaluationGateAgent, and LIVING_MEMORY.md. On Turn 1, it writes clean, minimal scaffolding and unit tests in seconds. The rubric gate runs locally in milliseconds, highlights any exact delta, and Turn 2 applies a surgical update. Fast generation paired with real compiler, test, and linter feedback consistently beats slow, unverified internal pondering.</li></ul><h4>3. Higher Efficiency: Eliminating Waste and Output Cost</h4><p>The cost breakdown reveals an equally striking reality: <strong>the loop is typically around one-seventh to one-eighth of the cost of the unharnessed single prompt</strong>.</p><p>Without context guardrails, the unharnessed model repeatedly duplicated schema definitions, wrote inline mock trivia data spanning hundreds of lines, and generated unrequested helper utilities that were never imported.</p><p>The harness prevents this bloat in three ways:</p><ol><li>harness_context.md explicitly enforces architectural boundaries (e.g. modular separation between the game engine and the trivia service).</li><li>The rubric’s <strong>Scope Discipline (Check 2.2)</strong> actively penalises extraneous abstractions. Because the agent knows it will be graded on tight adherence to scope, it writes concise, well-factored code rather than padding its output with defensive speculative fluff.</li><li><strong>Context Pruning Across Iterations (</strong><strong>include_contents=&quot;none&quot;):</strong> Rather than carrying the entire raw Turn 1 conversation history into Turn 2, the ADK builder agent prunes prior-turn chat history and receives only the distilled <strong>Iteration Evidence Receipt</strong> from LIVING_MEMORY.md (containing the score delta, pytest/ruff verification status, active failures, and next-turn remediation targets).</li></ol><h4>4. Resilience To Model Changes</h4><p>Finally, there is an overarching architectural benefit that transcends any individual run: <strong>resilience against model updates</strong>.</p><p>When teams rely on raw vibe coding, their production pipeline is fundamentally fragile. A prompt that works reasonably well on a Tuesday afternoon can suddenly fail on Wednesday when the provider deploys a minor model update, alters default temperature behaviour, or updates safety guardrails. When your only quality assurance is the model’s raw generative mood, you have no baseline.</p><p>A robust harness decouples and insulates us from changes to the model. The harness provides the immutable invariant rails: deterministic unit test runners, static analysers, type checkers, and semantic rubric gates. If a newer, smaller, or faster model stumbles on its initial attempt, the harness catches the defect instantly and guides autonomous remediation. You stop relying on model perfection and start relying on systemic verification.</p><h3>Off-the-Shelf Loop Mastery: The Loop Library and the loopy Skill</h3><p>Designing an effective developer loop from scratch can initially feel daunting. How do you calibrate the stopping conditions? How do you prevent an agent from running amok while refactoring a legacy codebase?</p><p>The good news is: <strong>you don’t have to invent every loop pattern from first principles.</strong></p><p>The agentic engineering community has begun codifying battle-tested loops into open-source catalogs. The preeminent reference in this space is <a href="https://signals.forwardfuture.com/loop-library/">The Loop Library</a> by Forward Future, accompanied by the open-source <a href="https://github.com/Forward-Future/loopy/tree/main/skills/loopy">loopy agent skill</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Sw8EEL3LpM3VEHOKDzXwEg.png" /><figcaption>The Loop Library</figcaption></figure><h4>What is The Loop Library?</h4><p>The Loop Library treats AI agent loops as <strong>bounded feedback systems with non-negotiable terminal states</strong>.</p><p>Every published loop in the library adheres to a rigorous architectural contract:</p><ol><li><strong>The Ready-to-Use Prompt:</strong> An unambiguous, declarative instruction that defines what the agent must do on each iteration.</li><li><strong>The Verify / Stop Criteria:</strong> Explicit, measurable terminal conditions (such as green CI or full test suite passage) that tell the loop when to stop.</li><li><strong>Strict Authority Boundaries:</strong> Clear operational tripwires separating read actions, code modifications, and irreversible operations (e.g. merging to main or releasing to production).</li></ol><h4>The loopy Agent Skill: Your In-IDE &quot;Loop Doctor&quot;</h4><p>Rather than manually browsing a website, you can install the <strong>loopy skill</strong> directly into your developer harness (such as Google Antigravity). When equipped with loopy, your agent becomes an intelligent loop architect:</p><ul><li><strong>Find &amp; Adapt:</strong> Ask your agent: <em>“Find a loop in the Loop Library to help me speed up my test suite.”</em> The agent queries the catalog, identifies the nearest matching pattern, and adapts its thresholds to your specific stack.</li><li><strong>Audit Existing Loops (“Loop Doctor”):</strong> Before letting an agent loose on a long-running task, ask loopy to audit your prompt. It inspects your instructions for weak checks, missing stopping conditions, or unbounded authority, hardening the loop before a single token is burned.</li><li><strong>Craft via Guided Interview:</strong> If no existing pattern fits your problem, loopy initiates a structured interview, probing you on the desired outcome, permitted tools, and exact success criteria to craft a bespoke, bounded loop.</li><li><strong>Save to </strong><strong>LOOPS.md:</strong> Once a loop proves effective in your repository, loopy can commit it directly into a version-controlled LOOPS.md file in the root of your project. This transforms successful loop patterns into shared, institutional team muscle memory.</li></ul><h4>Real-World Loop Patterns You Can Run Today</h4><p>Here are a few very cool, practical examples from <a href="https://signals.forwardfuture.com/loop-library/">The Loop Library</a> that illustrate how loop engineering transforms daily developer operations:</p><h4>1. The Code Housekeeper Loop</h4><ul><li><strong>The Problem:</strong> Technical debt, unformatted files, dead code, and low-priority linter warnings accumulate across repositories because human developers rarely have the time or patience to clean them up manually.</li><li><strong>How It Works:</strong> As defined in the <a href="https://signals.forwardfuture.com/loop-library/loops/housekeeper-loop/">Code Housekeeper Loop</a>, the agent wakes up, scans the codebase for static analysis warnings or dead imports, selects <strong>one single discrete issue</strong>, applies a targeted fix, runs the full test suite to prove zero regressions, and commits the change.</li><li><strong>Stopping Condition:</strong> Stops cleanly when all linter warnings are cleared or when a predefined maximum iteration cap is reached. Because it modifies only one file per turn, blast radius is completely contained.</li></ul><h4>2. The Fresh Clone Loop</h4><ul><li><strong>The Problem:</strong> A developer clones a repository, attempts to run the setup script, and immediately hits missing environment variables, unpinned dependency conflicts, or outdated documentation. “Works on my machine” strikes again.</li><li><strong>How It Works:</strong> The <a href="https://signals.forwardfuture.com/loop-library/loops/fresh-clone-loop/">Fresh Clone Loop</a> simulates the day-one developer experience in an isolated sandbox. It performs a completely clean clone, executes the documented bootstrap scripts, installs dependencies from scratch, and runs the initial test suite. If an installation step fails, it diagnoses the missing prerequisite, updates the setup.sh script or README.md, wipes the sandbox, and retries.</li><li><strong>Stopping Condition:</strong> The loop halts only when a fresh clone successfully builds, boots, and passes tests from an empty state without human assistance.</li></ul><h4>3. The Test Suite Speed Loop</h4><ul><li><strong>The Problem:</strong> Over time, test suites bloat. Slow database fixtures, redundant sleep calls, and un-parallelised test cases drag local feedback loops from thirty seconds to fifteen minutes, destroying developer velocity.</li><li><strong>How It Works:</strong> The <a href="https://signals.forwardfuture.com/loop-library/loops/test-suite-speed-loop/">Test Suite Speed Loop</a> profiles test execution times, identifies the slowest test files, isolates bottlenecks (such as redundant network calls or unmocked I/O), refactors the slowest test case, and verifies that the entire suite still passes with identical assertions.</li><li><strong>Stopping Condition:</strong> Halts when overall test suite execution time drops below a declared target threshold (e.g. under 60 seconds) with 100% test passage.</li></ul><h3>Summary &amp; What’s Next: Fuel for the Engine</h3><p>We have engineered an airtight structural harness (<a href="https://dev.to/gde/from-vibing-chaos-to-reliable-agents-the-what-why-and-how-of-harness-engineering-4ba0">Part 1</a>) and fired up an autonomous, self-evaluating execution loop with binary rubrics, tool trajectories, and the Ratchet Principle.</p><p><em>Now, what fuels the loop?</em></p><p>If you feed an autonomous loop vague conversational vibes and hand-waving, you won’t get software; you will get automated, looping chaos at 50 tokens a second.</p><p>In <strong>Part 3</strong>, we provide the ultimate fuel: <strong>Spec-Driven Development (SDD)</strong>. We will explore:</p><ul><li>Why code is disposable, but specifications are the permanent source of truth.</li><li>Authentic production spec blueprints (tasks/plan.md) with Three-Tier Operational Boundaries (<em>Always Do</em>, <em>Ask First</em>, <em>Never Do</em>).</li><li>Putting it all into practice with <strong>Google Antigravity</strong>: the unified harness across IDE, Extension, CLI, and SDK.</li><li>Native slash commands (/grill-me, /learn, /goal, /plan) vs formal SDD, and the <strong>4-Phase Hybrid Power Pattern</strong>.</li></ul><p>👉 <strong>Continue to Part 3: Spec-Driven Development For-the-Win: Disposable Code, Living Blueprints, and Building Like a Boss (Coming Soon)</strong></p><h3>Before You Go</h3><ul><li><strong>Please share</strong> 📢 this with anyone that you think will be interested. It might help them, and it really helps me!</li><li>Please <strong>give me loads of claps</strong>! 👏 (Just hold down the clap button.)</li><li>Please <strong>leave a comment</strong> 💬. Interaction is good!</li><li><strong>Add a star</strong> ⭐ on my repos!</li><li><strong>Follow</strong> 👉 and <strong>subscribe</strong> 🔔, so you don’t miss my content.</li></ul><h3>References and Useful Links</h3><h4>The Loop &amp; Autonomous Systems</h4><ul><li><a href="https://signals.forwardfuture.com/loop-library/">The Loop Library</a> &amp; <a href="https://github.com/Forward-Future/loopy/tree/main/skills/loopy">Loopy Skill</a></li><li><a href="https://www.rahuldhar.me/p/harness-engineering-vs-loop-engineering"><em>Harness Engineering vs. Loop Engineering</em> by Rahul Dhar (The Pragmatic Tech Lead)</a></li><li><a href="https://datasciencedojo.com/blog/harness-engineering-vs-loop-engineering/"><em>Harness engineering vs loop engineering</em> | Data Science Dojo</a></li><li><a href="https://www.mindstudio.ai/blog/loop-engineering-vs-harness-engineering"><em>Loop Engineering vs Harness Engineering: What’s the Difference and Which Do You Need?</em> | MindStudio</a></li><li><a href="https://flowtivity.ai/blog/loop-engineering-the-feedback-cycle-that-makes-ai-agents-work/"><em>Loop Engineering: The Feedback Cycle That Turns AI Agents Into Reliable Workers</em> | Flowtivity</a></li><li><a href="https://www.codecentric.de/en/knowledge-hub/blog/loop-harness-context-engineering-explained"><em>Loop, Harness, Context Engineering: The Terms Explained</em> | codecentric Knowledge Hub</a></li><li><a href="https://www.youtube.com/watch?v=ew6gBJNzC5w"><em>I’m a Senior Google AI PM. Here’s How I Build Loops (Finally, Loop Engineering Explained)</em> (Shubham Saboo on Product Faculty)</a></li></ul><h4>Dazbo’s Related Content</h4><ul><li><a href="https://github.com/derailed-dash/harness-engineering-demo"><em>Harness Engineering Demo Repository</em> on GitHub</a></li><li><a href="https://dev.to/gde/dialling-our-agents-to-11-agent-skills-you-need-to-be-using-2o2k"><em>Dialling Our Agents to 11: Agent Skills You Need to be Using!</em> by Dazbo</a></li><li><a href="https://dev.to/gde/dialling-our-agents-to-11-my-favourite-mcp-servers-3hbm"><em>Dialling Our Agents to 11: My Favourite MCP Servers</em> by Dazbo</a></li><li><a href="https://dev.to/gde/skills-sprawl-when-too-much-of-a-good-thing-confuses-your-ai-agent-4nij"><em>Skills Sprawl: When Too Much of a Good Thing Confuses Your AI Agent</em> by Dazbo</a></li><li><a href="https://dev.to/gde/meet-the-antigravity-extension-gemini-code-assist-on-steroids-in-whatever-ide-you-love-4p1n"><em>Meet the Antigravity Extension: Gemini Code Assist on Steroids, in Whatever IDE You Love</em> by Dazbo</a></li><li><a href="https://dev.to/gde/building-an-agentic-finops-platform-development-environment-setup-google-antigravity-mcps-and-4c43"><em>FinSavant Part 2: Building an Agentic FinOps Platform — Development Environment Setup, Google Antigravity, MCPs and Skills, and ADK Bootstrapping with Agents CLI</em> by Dazbo</a></li><li><a href="https://dazbo.co.uk"><em>Dazbo’s Portfolio</em></a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=95d3ac8f64cf" width="1" height="1" alt=""><hr><p><a href="https://medium.com/google-cloud/loop-engineering-in-the-harness-autonomous-self-correcting-systems-95d3ac8f64cf">Loop Engineering in the Harness: Autonomous, Self-Correcting Systems</a> was originally published in <a href="https://medium.com/google-cloud">Google Cloud - Community</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Vibing Chaos to Reliable Agents: The What, Why, and How of Harness Engineering]]></title>
            <link>https://medium.com/google-cloud/from-vibing-chaos-to-reliable-agents-the-what-why-and-how-of-harness-engineering-6e7e0443a4fc?source=rss-58a0bd8ba726------2</link>
            <guid isPermaLink="false">https://medium.com/p/6e7e0443a4fc</guid>
            <category><![CDATA[context-engineering]]></category>
            <category><![CDATA[agent-skills]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[mcps]]></category>
            <category><![CDATA[harness-engineering]]></category>
            <dc:creator><![CDATA[Dazbo (Darren Lester)]]></dc:creator>
            <pubDate>Mon, 28 Sep 2026 09:59:24 GMT</pubDate>
            <atom:updated>2026-10-26T12:12:26.947Z</atom:updated>
            <content:encoded><![CDATA[<p>You are in Part 1 of the series <strong><em>“Beyond Vibe Coding: The Engineering Blueprint for Reliable AI Agents”</em></strong></p><h3><strong><em>🗺️ The Series Roadmap: Beyond Vibe Coding</em></strong></h3><p><em>Agentic software development is drowning in fragmented buzzwords: Harness Engineering, Loop Engineering, Context Engineering, Spec-Driven Development… Are they competing frameworks? Redundant techniques? Or fundamentally different and essential?</em></p><p><em>This three-part series unifies these concepts into a single, cohesive architectural blueprint for building software with AI without writing code manually:</em></p><ul><li><strong><em>Part 1: From Vibing Chaos to Reliable Agents: The What, Why, and How of Harness Engineering (This Article)</em></strong><em><br>We diagnose the issues of vibe coding, unpack the paradigm shift of shifting developer interventions left, establish our bedrock formula (</em><strong><em>Agent = Model + Harness</em></strong><em>), explore a motorsport analogy, and deconstruct deterministic foundations that make up a production developer environment.</em></li><li><a href="https://medium.com/google-cloud/loop-engineering-in-the-harness-autonomous-self-correcting-systems-95d3ac8f64cf?postPublishedType=repub"><strong><em>Part 2: Loop Engineering in the Harness: Autonomous, Self-Correcting Systems</em></strong></a><em><br>We fire up the dynamic runtime engine of the harness: deconstructing the loop components, establishing hybrid evaluation rubrics (deterministic tests + LLM goal fidelity + tool trajectory verification), and applying the Ratchet Principle.</em></li><li><strong><em>Part 3: Spec-Driven Development For-the-Win: Disposable Code, Living Blueprints, and Building Like a Boss (Coming Soon)</em></strong><em><br>We explore Spec-Driven Development (where code is disposable and specifications are the real IP), master Google Antigravity’s unified developer harness and interactive slash commands (</em><em>/grill-me, </em><em>/plan, </em><em>/goal, </em><em>/boost, </em><em>/learn), and combine formal specs with tactical workflows via the 4-Phase Hybrid Power Pattern.</em></li><li><strong><em>Part 4: Scaling the Harness: Practical Loop Patterns, ADK 2 Graphs, Recursive Harness², and Harness Maintenance (Coming Soon)</em></strong><em><br>We bring everything together with practical loop patterns you can wire into any repo (from declarative Markdown and </em><em>loopy to Google ADK 2 graphs), keep context clean across iterations with short-lived subagents, explore Google Research’s recursive Harness</em><strong><em>²</em></strong><em> framework, and cover ongoing harness maintenance.</em></li></ul><h3>Introduction: The Chaos of Vibe Coding</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hlo6rizVz8VS9Py0hGHMGw.png" /><figcaption>From Vibing Chaos to Harness Engineeering</figcaption></figure><p>Was your first vibe coding experience like mine? Type in a single prompt and then watch an entire application spring into existence. It felt like wizardry.</p><p>It was clear that software development had fundamentally changed. We had entered the era of “vibe coding” — that intoxicating workflow where you lean back in your chair, sip your brew, prompt an LLM with a breezy idea, and casually mash the approval button while the model writes hundreds of lines of code across half a dozen files.</p><p>For quick hobby prototypes, weekend experiments, or throwing together an interactive mock-up to impress a client on a Friday afternoon, vibe coding is both effective and <em>brilliant fun</em>. You watch a terminal stream lines of code at lightspeed, and for a fleeting moment you feel like an omnipotent 100x developer. You can now build things in hours that would have once taken a competent coder months.</p><p><em>Then Monday morning arrives.</em></p><p>When you attempt to take your vibe-coded solution anywhere near a real production environment, reality hits you like a cold, wet Monday in London. What felt like intoxicating velocity on Friday turns into an architectural crime scene by Monday. The code might run, but beneath the surface lies a tangled web of subtle bugs, orphaned functions, unverifiable assumptions, and <em>utter nonsense</em>.</p><p>You quickly uncover the dirty secret behind unconstrained vibing:</p><ul><li><strong>Hallucination City:</strong> The model confidently invents deprecated API methods and fictitious URLs, synthesises nonexistent package dependencies, and hallucinates framework features that simply don’t exist. Once an agent hallucinates a dependency, every subsequent turn doubles-down on the fiction, inventing bizarre workarounds for libraries that were never real in the first place.</li><li><strong>Alert &amp; “LGTM” Fatigue:</strong> After reading your forty-seventh sprawling diff in two hours, your eyes glaze over. You find yourself mindlessly clicking “Approve” or typing “LGTM” without actually understanding or checking the changes, silently hoping the AI knows what it’s doing. (Spoiler: <em>it doesn’t.</em> You haven’t eliminated manual labour; you have merely swapped creative software engineering for bleary-eyed cursory proofreading.)</li><li><strong>Solutions That Don’t Do What You Wanted:</strong> The code runs without syntax errors, but the actual business logic subtly diverges from what you originally intended. Maybe because you didn’t really know what you wanted in the first place, and you just hope that the model has magically filled in all the blanks in your fluffy request.</li><li><strong>Unmaintainable, Inscrutable Code:</strong> The agent generates dense blocks of code without architectural coherence. It stitches together three different design patterns across four files. You don’t understand how it works and you haven’t really checked it. And you’d probably never reveal the code to your competent peers. <em>“What if they actually look at this?”</em> Your code was tech-debt on day one.</li><li><strong>No Documentation:</strong> Zero README updates, no Architecture Decision Records (ADRs), no API contracts, and no inline explanations of why specific trade-offs were made. The key knowledge that maybe lives in your head is completely absent from the documentation.</li><li><strong>Insecure &amp; Risky:</strong> Hardcoded secrets, wide-open IAM permissions, complete absence of rate limits, and unescaped SQL queries slipped into the codebase because the agent prioritised making the code run over making it safe. Maybe you didn’t like a warning in a log and you told the agent <em>“Make this warning disappear.”</em> The agent said: <em>“Hold my beer.”</em></li><li><strong>Inefficient Token Use &amp; The Manual Loop Furnace:</strong> You find yourself trapped in a soul-crushing manual loop: copy terminal error → paste into chat → get half-baked fix → copy next error → watch the context window bloat → watch your token bill skyrocket. You aren’t architecting software; you are acting as an organic clipboard between a compiler and an LLM. Inevitably, unless you happen to be paying for a premium model subscription, you’re going to hit your model quota.</li><li><strong>Lack of Scalability &amp; Portability:</strong> The code works on your machine, under your specific node or python version, with zero containerisation or repeatable environment configuration. Hand it to a colleague or push it to CI, and it instantly shatters. <em>Onboarding friction</em>, anyone?</li><li><strong>Purpose Drift:</strong> By turn eight of a messy conversation, the model has completely forgotten the initial rules and constraints you laid out in turn one. (Assuming you specified any constraints at all.) It begins re-inventing components it already built or quietly undoing your earlier bug fixes.</li></ul><p>The conclusion is inescapable: <strong>you cannot ship vibes to production.</strong></p><p>Yet, none of us wants to retreat to typing out thousands of lines of boilerplate by hand. The goal remains what it has always been: <strong>to create robust, scalable software solutions without writing code manually</strong>.</p><p>In an insightful <a href="https://www.youtube.com/watch?v=F8EZJAm9iO8">Google Cloud Tech interview</a> with Tilde Thurium, <strong>Ryan Lopopolo</strong> (Principal Engineer, Agentic Cloud Platform @ Google, who first coined the term <em>“agent harness”</em>) laid out his own radical ambition: to never write code by hand again. As Ryan wryly admitted: <em>“I aspire to be a lazy programmer.”</em> Don’t we all?</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fsiteproxy.ruqli.workers.dev%3A443%2Fhttps%2Fwww.youtube.com%2Fembed%2FF8EZJAm9iO8%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fsiteproxy.ruqli.workers.dev%3A443%2Fhttps%2Fwww.youtube.com%2Fwatch%3Fv%3DF8EZJAm9iO8&amp;image=https%3A%2F%2Fsiteproxy.ruqli.workers.dev%3A443%2Fhttps%2Fi.ytimg.com%2Fvi%2FF8EZJAm9iO8%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/f5718b3123fdc7b418b92ec661d25c0b/href">https://medium.com/media/f5718b3123fdc7b418b92ec661d25c0b/href</a></iframe><p>Having demonstrated that teams can build and ship massive production systems with zero manual typing by establishing that <em>“humans steer, agents execute”</em>, Lopopolo proved that hands-off development is genuinely achievable.</p><p>The answer isn’t to abandon AI agents or wait for newer models; it is to fundamentally change how we construct, control, and constrain them.</p><h3>The Paradigm Shift: Shifting Developers Left</h3><p>To understand why unconstrained vibe coding collapses, we should look at <em>where</em> developers are forced to spend their energy.</p><p>In naive AI interactions, developer intervention happens almost entirely <strong>on the right</strong>. That is: downstream, after the code has already been generated. You spend your afternoon scrolling through hundreds of lines of inscrutable diffs, tweaking chat prompts to fix syntax quirks, or rewriting hallucinations. You are perpetually reacting to failure after the fact, acting as an exhausted organic filter for broken output.</p><p>Harness engineering <strong>shifts developer intervention to the left</strong>.</p><p>Instead of manually reviewing broken outputs or wrestling with repetitive errors downstream, you invest your engineering effort upstream <em>before</em> the agent ever runs. You engineer the environment: defining rules, embedding automated linters and type-checkers, configuring deterministic sandboxes, and establishing clear evaluation rubrics. When the agent operates within an engineered environment, failures are caught and corrected autonomously inside the execution loop, long before a human ever inspects the resulting pull request.</p><p>Enter <strong>Harness Engineering</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PqoW0VpawX7JXmmiM_g84Q.png" /><figcaption>The Paradigm Shift</figcaption></figure><h3>If It’s Not the Model, It’s the Harness</h3><p>To understand what <em>harness engineering</em> is, you must start with this fundamental equation:</p><p><strong>Agent = Model + Harness</strong></p><p>And with this comes the ultimate maxim of modern AI development:</p><blockquote><strong>“If it’s not the model, it’s the harness.”</strong></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OZfYvXiQBNtAmFIPOgao7A.png" /><figcaption>The Agent: Model + Harness</figcaption></figure><p>We have been conditioned by tech headlines to obsess over frontier models. Every week brings another benchmark, another parameter increase, and another claim of AGI just around the corner. But here is the hard reality: <strong>the harness is arguably more important than the model.</strong></p><p>An okay, cost-effective model housed within a robust harness will consistently outperform a bleeding-edge frontier model with a poor harness. Why? Because the model is inherently probabilistic, whereas serious software engineering demands determinism. The harness is what bridges that chasm. It supplies the missing structure: living memory, automated validation, static analysis, type checking, sandboxing, and execution guardrails.</p><p>Think of it like high-performance motorsport:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hq5IrxLgF07TkhwainWGeg.png" /><figcaption>The Motorsport Analogy for Agents and the Harness</figcaption></figure><ul><li><strong>The Model is the Driver:</strong> Incredibly talented, capable of split-second reasoning and creative problem-solving, but prone to fatigue, distraction, and occasional moments of reckless overconfidence. Left entirely to their own devices on an open airfield, they might pull off some breathtaking doughnuts, but they can’t win a grand prix.</li><li><strong>The Entire Race Car IS the Harness:</strong> The complete, precision-engineered machine built to translate the driver’s intent into track-winning performance:</li><li><strong>The Engine &amp; Transmission (Loop Engineering):</strong> The internal dynamic execution engine of the harness. It delivers torque to the wheels through a disciplined cycle of intake, combustion, and telemetry. Iterative execution, continuous evaluation, and immediate feedback cycles.</li><li><strong>The Chassis (Structural Harness):</strong> The living rules (AGENTS.md / GEMINI.md), directory layouts, and sandboxed runtimes that keep the car glued to the tarmac, dictating the exact physical limits of where the vehicle can steer.</li><li><strong>The Brakes &amp; Pit Telemetry (Safety &amp; Evaluation Gates):</strong> The deterministic test suites and escalation protocols that prevent the driver from wrapping the machine around a barrier at 180 mph.</li></ul><p>You wouldn’t put Lewis Hamilton behind the wheel of a V12 engine block with no chassis or brakes and expect him to finish a single lap at Silverstone, let alone win the race. So why do the same with agentic software development and expect robust results?</p><h4><strong><em>Dazbo’s Battle Scar: The Model Upgrade Fallacy</em></strong></h4><blockquote><em>Whenever an agentic workflow breaks down in development — falling into repetitive loops, introducing subtle runtime bugs, or failing to follow architectural patterns — the instinctive developer reaction is: </em>“We need a smarter model. Let’s switch to the latest 2-trillion parameter frontier model.”</blockquote><blockquote><em>Almost every time, this is a fallacy. Upgrading the model without improving your harness simply means your agent will generate more articulate, harder-to-spot hallucinations at four times the token cost. If your agent is failing, don’t reach for a bigger model first. Ask yourself: </em>Did the harness enforce a test? Did the harness provide the exact schema? Did the harness restate the goal?<em> If it’s not the model, it’s the harness!</em></blockquote><p>Crucially, <strong>the harness is model-agnostic</strong>. Whether you are routing requests through a cost-effective, high-velocity workhorse like gemini-3.8-flash, or a deep-reasoning powerhouse, a well-engineered harness remains stable, repeatable, and robust. You can hot-swap the model under the hood without having to redesign your entire engineering lifecycle.</p><h3>The Model is a Commodity; The Harness is the IP</h3><p>This reveals a fundamental commercial reality of AI software engineering. As my friend and fellow Google Developer Expert <strong>Jaroslav Pantsjoha (JP)</strong> highlighted in <a href="https://medium.com/google-cloud/the-ai-native-developer-experience-part-2-harness-engineering-with-google-antigravity-7fb72dab243f">his analysis of the AI-native developer experience</a>: <strong>the model is a commodity; the harness is where your true competitive advantage and intellectual property reside.</strong></p><p>Every organisation has access to the exact same frontier foundational models via public APIs. Simply renting access to raw model intelligence provides zero lasting competitive advantage. <strong>Your true intellectual property and your team’s operational velocity live entirely inside your harness</strong>:</p><ol><li><strong>Your living rules and architectural guardrails</strong> that encode decades of hard-won domain knowledge.</li><li><strong>Your curated suite of tool integrations and MCP servers</strong> that securely bridge AI agents into your private internal databases and microservices.</li><li><strong>Your deterministic evaluation rubrics and automated test harnesses</strong> that guarantee safety and compliance before any code reaches production.</li><li><strong>Your version-controlled living specifications</strong> that turn disposable AI code into maintained, documented, and resilient software assets.</li></ol><p>Foundational models will continue to be commoditised, upgraded, and replaced. But a well-architected enterprise harness compounds in value with every single engineering iteration.</p><h3>Untangling the Triad: Context, Harness, and Loop Engineering</h3><p>Before we deconstruct the components, it is worth clarifying how three terms frequently heard across the industry map directly back to our bedrock formula (<strong>Agent = Model + Harness</strong>):</p><ul><li><strong>Context Engineering:</strong> Determines what the agent <strong>knows</strong>. It is the cognitive payload the harness feeds to the model, including hierarchical rules (AGENTS.md or GEMINI.md that can be global, team, domain, or project-specific), active retrieval (RAG), agent skills, and other sources of knowledge.</li><li><strong>Structural Harness Engineering:</strong> Determines what the agent <strong>can touch and where it runs</strong>. It provides the deterministic environment: tools, APIs, MCP servers, sandboxed execution runtimes, security guardrails, living file state, and blast-radius (damage control).</li><li><strong>Loop Engineering:</strong> Determines how the agent <strong>iterates and converges</strong> over time. It is <strong>the dynamic runtime engine of the harness</strong>; the autonomous process of observing, acting, judging against rubrics, recording memory, and knowing when to stop.</li></ul><p>Crucially, <strong>they are not three competing systems; loop and context are the internal functional dimensions of the harness itself.</strong> If context is the agent’s briefing pack and the loop is its working cadence, the harness is the complete laboratory workbench in which all of it takes place.</p><p>And we can see that state (such as markdown files) plays a significant role in all three.</p><h3>Deconstructing the Harness: The Deterministic Foundations</h3><p>If the model represents the non-deterministic, probabilistic core of our system, the harness represents everything else: the <strong>deterministic components </strong>that guide, restrain, and validate every single action the model takes.</p><p>So, what actually lives inside a robust developer harness?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9uxOW_lQ1kX_gPuboxm0FA.png" /><figcaption>The Foundations of the Developer Harness</figcaption></figure><p>Let’s break down the essential layers:</p><h4>1. Hierarchical Context &amp; Living Rules: “De-Slopping” by Design</h4><p>Your agent should never have to guess your engineering standards, architectural decisions, or toolchain preferences. In an ad-hoc vibe coding setup, you might type: <em>“Please write clean Python and test it.”</em> The model nods, writes an unformatted script, tests it with a couple of inline assert statements, and calls it a day.</p><p>In her analysis <a href="https://www.builder.io/blog/de-slop-ai-generated-codebase"><em>How to De-Slop an AI-Generated Codebase</em></a>, Alice Moore highlights why telling an AI to <em>“clean this up”</em> or <em>“write clean code”</em> inevitably fails: the instruction is too vague. AI “slop” can only be prevented by translating human taste into <strong>concrete, non-negotiable rules</strong>.</p><p>In an engineered harness, we establish our context in files like this:</p><ul><li><strong>Global Configuration (</strong><strong>GEMINI.md or </strong><strong>AGENTS.md):</strong> Cross-project standards including language preferences, mandatory package managers (e.g. using uv for Python), security baseline rules, and zero-tolerance policies for inline execution hacks.</li><li><strong>Workspace-Specific Context (</strong><strong>my-project/GEMINI.md):</strong> Architecture blueprints, directory layouts, data sources, and project goals.</li><li><strong>Automated Guardrails:</strong> Explicit rules that we must always adhere to. For example, performing static analysis after each change. Or, for example, requiring the agent to practice strict Test-Driven Development (TDD). I.e. writing the failing test before implementing code.</li><li><strong>Documentation Rules:</strong> Mandatory upkeep of core documentation (READMEs, Architecture Decision Records, product specs, task lists) using standardised workflows so the codebase never drifts into an undocumented quagmire.</li></ul><p>To see what this looks like in practice, here is an authentic snippet from my own global GEMINI.md file:</p><pre># Agent Execution Guardrails &amp; Standards<br><br>## Python Guidance &amp; Tooling<br><br>- **Environment Source of Truth**: Respect the Python version defined in<br>  `pyproject.toml`. For new greenfield projects, default to `python 3.13`.<br>- `uv` is the preferred tool for Python package and environment management.<br>- **Avoid Inline `python -c` Commands**: Do NOT execute inline Python snippets<br>  using `python -c &quot;...&quot;` or `python3 -c &quot;...&quot;`. Because Antigravity cannot <br>  safely prefix-match arbitrary inline code payloads, this triggers a manual <br>  approval prompt every time. Instead, write one-off helper logic to a <br>  scratch script in your scratch directory and execute it cleanly via <br>  `uv run python scratch/helper.py`.<br>- Always include top-of-module docstrings that explain what a given .py <br>  file does, why it exists (i.e. the value it provides) and an outline of <br>  how it works.<br>- Use rich (`pip install rich`) when building console / CLI applications.<br>- **Zero-Tolerance for Suppression Hacks**: Never bypass linter or <br>  type-checker errors using `# noqa`, `# type: ignore`, or `@ts-ignore` <br>  comments. Resolve the underlying type contract or architectural root cause.<br>- When adding features or making changes that are anything beyond trivial, <br>  **use TDD**. Create the test, verify it fails, make the change, and then verify <br>  the test passes.<br>- To ensure Python code adheres to required standards, the following <br>  commands **must** be run after creating or modifying any `.py` files:<br>  ```bash<br>  uvx codespell@latest -s        # check spelling and show summary<br>  uvx ruff@latest check --fix .  # perform static checks and auto-fix<br>  ```<br><br>## Agent and AI Development<br><br>- NEVER try to downgrade a Gemini model in code or variables. E.g. NEVER <br>  try to change `gemini-3.8-flash` to `gemini-3-flash`.<br>- Unless otherwise instructed, use the latest generally available Gemini <br>  models available; always check what the latest is. This also goes for <br>  image generation.<br>- Unless I say otherwise, agents should be built using the Google ADK <br>  (`google-adk`) and the Google Gen AI (`google-genai`) packages. <br>- AVOID using the `google-generativeai` package. This is deprecated in <br>  favour of the newer GenAI `google-genai` SDK.</pre><p>We’re not politely pleading with the model. We are creating rigid rails.</p><p>The rule banning inline python -c snippets is a classic harness engineering battle scar. In a sandboxed developer environment, the execution engine auto-approves safe binary prefixes (such as uv run python ...). When an unconstrained model starts generating dynamic inline code strings, the runtime cannot prefix-match the payload and is forced to interrupt the developer with a manual approval prompt every time it makes a minor tweak to the Python it wants to execute… Every ten seconds!</p><p>Similarly, the rule forbidding # type: ignore and # noqa suppressions closes a notorious AI loophole. When confronted with strict linting or typing rules, an unsupervised LLM will frequently take the path of least resistance by slapping suppression comments across the file to silence the compiler. By outlawing suppressions in the harness rules, the model is compelled to fix the actual defect.</p><h4>2. Living State &amp; Version-Controlled Memory</h4><p>Human software engineers don’t attempt to hold an entire multi-week project in their brains; we maintain backlog boards, write technical specifications, document decisions, and keep scratchpads.</p><p>Yet developers routinely expect an AI model to hold twenty architectural considerations across a forty-turn chat window. The moment the conversation grows long, the context window compacts, and the agent quietly forgets earlier decisions. And let’s not forget the <em>“lost-in-the-middle”</em> problem that we frequently observe in models.</p><p>The harness solves this by providing agents with structured, file-backed <strong>living state</strong>. Although there are no rigid rules for <em>HOW</em> you structure this, here are some suggestions from me:</p><p><strong>Strategic &amp; Architectural Truth (The Blueprints):</strong></p><ul><li>PRODUCT.md: Defines the business “why”, user personas and user stories, core problem statement, and explicit non-goals (what the system is deliberately <em>not</em> building, preventing the agent from over-engineering features you never asked for).</li><li>SPEC.md: Machine-readable functional requirements, API contracts, and behavioural scenarios (Given / When / Then).</li><li>ARCHITECTURE.md: System topology, component boundaries, database schemas, data flow and design decisions.</li><li>DESIGN.md: UI design, styling rules, token definitions (colours, typography, spacing), and accessibility constraints; we can even export or synchronise this file with tools like Google Stitch via MCP.</li></ul><p><strong>Operational &amp; Execution State (The Working Memory):</strong></p><ul><li>TODO.md (or active backlog): Real-time checklists where tasks are marked off as completed.</li><li>PLAN.md: Phased vertical slices and dependencies mapped out for the immediate objective.</li></ul><p><strong>Transient Execution State (The Scratchpad):</strong></p><ul><li>scratch/: Ephemeral execution notes, raw compiler traces, and intermediate data files that persist across model turns without polluting Git history. You would exclude this folder in your .gitignore.</li></ul><p><em>(We will explore how to author, stress-test, and execute these living blueprints in depth in Part 3: Spec-Driven Development For-the-Win. Coming soon!)</em></p><p>Because these files are stored in standard Markdown directly in the workspace, <strong>they can be version-controlled with Git and shared seamlessly across developer sessions and different machines</strong>. If an agent session crashes, if your laptop runs out of battery, or if you hand the task over to a colleague across the globe, the harness preserves the living state. The agent simply reads the task list on turn one and resumes exactly where it left off.</p><h4>3. Agent Tools: Grounding Models in the Real World</h4><p>Language models in a vacuum are just predictive text generators; they interact with the outside world through tools. A robust harness equips the agent with a multi-layered execution surface:</p><ul><li><strong>Native Functions:</strong> Strongly typed functions in the code, for fast deterministic operations.</li><li><strong>Sandboxed Shell and CLIs:</strong> Compilers, linters (ruff, eslint), test runners (pytest), package managers (uv, npm), and cloud CLIs (e.g. gcloud) executed inside isolated shell environments.</li><li><strong>The Model Context Protocol (MCP):</strong> A standard that lets agents discover and execute tools without writing bespoke glue code for every service.</li></ul><h4><strong><em>Dazbo’s Battle Scar: MCP Sprawl &amp; Tool Collision</em></strong></h4><blockquote><em>When developers first discover MCP, the temptation is to connect twenty different servers simultaneously: GitHub, Slack, Postgres, BigQuery, GKE, Firestore, Stitch, Confluence, Jira, Cloud Logging…</em></blockquote><blockquote><em>Don’t do it. Connecting 20 MCP servers could easily dump 150+ tool schemas into your model’s context. That’s a lot of wasted tokens before you’ve even typed an instruction. Worse, models experience </em>tool choice paralysis<em> and routing collisions. As I detailed in </em><a href="https://dev.to/gde/dialling-our-agents-to-11-my-favourite-mcp-servers-3hbm">Dialling Our Agents to 11: My Favourite MCP Servers</a><em>, keep your active MCP server suite lean, curated, and strictly relevant to the task at hand.</em></blockquote><h4>4. On-Demand Context: Progressive Disclosure via Agent Skills</h4><p>Flooding an LLM’s context window with thousands of lines of reference manuals on every turn is a great way to confuse your model and increase your token costs.</p><p>Instead, the harness can use <em>agent skills</em>, which use a mechanism called <strong>progressive disclosure</strong> <strong>to load on-demand</strong>. Progressive disclosure operates in three levels:</p><ul><li><strong>Level 1 (Metadata):</strong> At startup, the agent only reads tiny YAML frontmatter headers (e.g. name and trigger description, typically consuming under 100 tokens).</li><li><strong>Level 2 (Instructions):</strong> Only when a relevant task is initiated does the agent dynamically load the full SKILL.md instruction set.</li><li><strong>Level 3 (Resources):</strong> Deep reference scripts, schemas, and templates are only pulled into context when needed.</li></ul><p>Your agent stays lightning-fast, burns minimal tokens, and possesses instant access to all the knowledge and operational expertise it needs.</p><p>Feels a bit like cheating, doesn’t it?</p><h4><strong><em>Dazbo’s Battle Scar: Skills Sprawl &amp; Context Pollution</em></strong></h4><blockquote><em>If skills are so effective, why not install hundreds of them?</em></blockquote><blockquote><em>As I explored in </em><a href="https://dev.to/gde/skills-sprawl-when-too-much-of-a-good-thing-confuses-your-ai-agent-4nij">Skills Sprawl: When Too Much of a Good Thing Confuses Your AI Agent</a><em>, skills hoarding is a trap.</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/661/1*RSdmHxDQqGuI_YT30x7ggA.png" /><figcaption>Too many skills — It’s a trap!</figcaption></figure><blockquote><em>Even though Level 1 progressive disclosure reads only the YAML frontmatter, loading 150+ skills can easily inject over 15,000 tokens and routing overhead into every single model turn.</em></blockquote><blockquote><em>The result? Severe LLM decision fatigue, tool selection degradation, and skill routing collisions. If your agent is building a Python cloud backend, it shouldn’t be wading through metadata for React video rendering or mobile UI frameworks. Curate your active skills, prune what you don’t use, and disable non-essential sub-skills in your configuration. A disciplined harness is a clean harness. (For a deep dive into the top skills I rely on daily, check out </em><a href="https://dev.to/gde/dialling-our-agents-to-11-agent-skills-you-need-to-be-using-2o2k">Dialling Our Agents to 11: Agent Skills You Need to be Using!</a><em>.)</em></blockquote><h4>5. Packaging Knowledge &amp; Capability: The Rules, Skills, and MCP Triad</h4><p>Once you begin equipping your harness with these deterministic components, an obvious engineering question arises: <em>How do you manage all this?</em></p><p>In practice, we organise with three complementary layers:</p><ol><li><strong>Rules (The Guardrails &amp; Standards):</strong> Standing instructions that ensure compliance (e.g. safe command syntax, directory layouts, and security policies).</li><li><strong>MCP Servers (The Tools &amp; Senses):</strong> Execution interfaces that connect the agent to external infrastructure, databases, and APIs.</li><li><strong>Agent Skills (The Procedural Playbooks):</strong> On-demand workflows and recipes that teach the agent <em>how</em> and <em>when</em> to use those tools effectively.</li></ol><p>While open protocols like MCP have standardised the tool layer, there is not yet a single universal industry standard for how the entire capability stack should be packaged.</p><p>However, an increasingly popular architectural pattern — supported natively by developer harnesses like <strong>Google Antigravity</strong> — is the <strong>Plugin</strong>. Rather than forcing developers to manually wire up an MCP server, write separate prompt rules, and install five standalone skills, a plugin acts as an aggregated delivery vehicle:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2MsnlC-xAYraWbuLafwa9Q.png" /><figcaption>The Plugin</figcaption></figure><p>A prime example of this pattern in the Google ecosystem is the <strong>google-cloud-developer plugin</strong>, which looks like this:</p><ul><li><strong>Packaging Manifest (</strong><strong>plugin.json):</strong> The schema descriptor defining metadata, versioning, and dependencies.</li><li><strong>Standing Safety Rules (</strong><strong>rules/*.md):</strong> Markdown rules (like google-cloud-discovery.md) that establish pre-flight command validation and guardrails.</li><li><strong>Modular Agent Skills (</strong><strong>skills/*/SKILL.md):</strong> Individual skill directories with YAML frontmatter headers and procedural Markdown instructions—such as google-cloud-recipe-auth, gcloud, and dynamic discovery (finding-google-skills).</li><li><strong>Tool Connectivity (</strong><strong>mcp.json):</strong> A JSON configuration declaring active MCP endpoints—such as the streamable-HTTP connection to Google&#39;s official developer-knowledge server.</li></ul><pre>google-cloud-developer/<br>├── plugin.json                    # Plugin metadata &amp; agent-plugins.org schema<br>├── mcp.json                       # MCP server connection endpoints<br>├── rules/<br>│   └── google-cloud-discovery.md  # Standing guardrails &amp; catalog rules<br>└── skills/<br>    ├── finding-google-skills/<br>    │   └── SKILL.md               # Dynamic on-demand catalog discovery<br>    ├── gcloud/<br>    │   └── SKILL.md               # Gcloud CLI validation &amp; execution<br>    └── google-cloud-recipe-auth/<br>        └── SKILL.md               # Authentication recipes</pre><p>Whether packaged as composite plugins or managed as standalone assets, the goal is identical: <strong>providing agents with both the architectural rules and the deep domain knowledge required to execute safely.</strong></p><h4>My Curated Toolkit</h4><p>In my own daily engineering workflow, I rely on a focused blend of plugins and standalone skills, which include (but are not limited to):</p><ul><li><a href="https://github.com/google/skills/tree/main/plugins/cloud/google-cloud-developer"><strong>google-cloud-developer plugin</strong></a><strong>:</strong> Deep GCP architectural rules, safe execution guardrails, and native developer knowledge retrieval.</li><li><a href="https://github.com/google/agents-cli/tree/main/skills"><strong>Google Agents CLI</strong></a><strong>:</strong> Coding best practices, scaffolding, and deployment blueprints for building AI agents with the Google Agent Development Kit.</li><li><a href="https://github.com/google/skills/tree/main/skills/cloud/agent-platform-eval-flywheel"><strong>agent-platform-eval-flywheel</strong></a><strong>:</strong> Google Cloud&#39;s evaluation flywheel methodology for measuring agent quality, LLM-as-judge scoring, and regression datasets.</li><li><a href="https://github.com/addyosmani/agent-skills/tree/main/skills/spec-driven-development"><strong>spec-driven-development</strong></a><strong>:</strong> Bridges intent to code by generating machine-executable specifications and behavioural contracts before coding begins (the focus of Part 3).</li><li><a href="https://github.com/addyosmani/agent-skills/tree/main/skills/test-driven-development"><strong>test-driven-development</strong></a><strong>:</strong> Enforces strict red-green-refactor loops, guaranteeing code isn&#39;t accepted until automated test suites pass.</li><li><a href="https://github.com/addyosmani/agent-skills/tree/main/skills/code-review-and-quality"><strong>code-review-and-quality</strong></a><strong>:</strong> Automated multi-axis code reviews evaluating maintainability, security, and simplicity before changes enter the main branch.</li><li><a href="https://github.com/vercel-labs/skills"><strong>Vercel’s </strong></a><a href="https://github.com/vercel-labs/skills"><strong>find-skills</strong></a><strong>:</strong> Dynamic skill discovery, allowing the agent to find and fetch installable skills from public registries on-demand without manual preloading.</li><li><a href="https://github.com/derailed-dash/dazbo-agent-skills/tree/main/skills/organise-agent-skills"><strong>organise-agent-skills</strong></a><strong>:</strong> My own skill for auditing installed skills, pruning token overhead, and managing on-demand progressive disclosure to prevent skills sprawl.</li><li><a href="https://github.com/derailed-dash/dazbo-agent-skills/tree/main/skills/maintaining-core-documentation"><strong>maintaining-core-documentation</strong></a><strong>:</strong> My own agent skill for synchronising READMEs, architectural docs, and task lists whenever code changes.</li></ul><h4>6. Sandboxing &amp; Blast-Radius Containment</h4><p>A robust harness never allows an agent unfettered access to host operating systems or production networks. Through containerised agent sandboxes, filesystem isolation, and execution policies, the harness establishes hard boundaries.</p><p>If an agent attempts an unreviewed destructive action — such as dropping a database table, executing an unsandboxed shell command that alters host configuration, or modifying .git/ internal objects — the harness halts execution and demands explicit human verification. Sandboxing transforms autonomous agents from dangerous mavericks into safe, bounded wingmen. (See what I did there?)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/737/0*lgZkiT8-gAZCue74.jpg" /><figcaption>Maverick — can be your wingman</figcaption></figure><h4>7. Observability &amp; Trajectory Monitoring: Beyond Traditional Logging</h4><p>In traditional software engineering, when a service fails, we inspect a deterministic stack trace and a linear log file: line 42 threw a NullPointerException. Or we run our unit tests and see exactly what test failed.</p><p>Autonomous AI agents don’t work like that. An agent doesn’t execute a single static function; it embarks on a multi-step journey toward a goal. In AI engineering, that journey is known as the agent’s <strong>trajectory</strong>.</p><blockquote><strong>What is an Agent Trajectory?</strong><br>An agent trajectory is the complete, chronological sequence of intermediate states, reasoning steps, tool calls, environment responses, and course corrections the model takes from its initial prompt to the final output.</blockquote><p>Think of it as a state machine where every step consists of a continuous route.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GJJdqWUPc24sfenzB5i07A.png" /><figcaption>Agent Trajectory</figcaption></figure><p>For example, a four-turn trajectory might look like:<br>Turn 1: Read SPEC.md → Turn 2: Run pytest (fails) → Turn 3: Edit code → Turn 4: Re-run test (passes).</p><p>When an agent fails, it rarely crashes with a fatal error. Instead, it suffers from <strong>trajectory drift</strong>:</p><ul><li>Did it misinterpret a requirement back on turn 2?</li><li>Did a sandboxed CLI return an ambiguous warning on turn 4 that poisoned the agent’s reasoning on turn 5?</li><li>Did the model get stuck in a “thrashing loop” between turns 8 and 12, fruitlessly editing the same file back and forth?</li><li>Did it choose a blunt command-line tool when an exact MCP endpoint was sitting right in its prompt?</li></ul><p>Without trajectory observability, you are left staring at a vague final apology from the model, completely blind to where the journey went off the rails.</p><p>A robust developer harness treats trajectory monitoring as a first-class citizen:</p><ol><li><strong>Step-Level Execution Traces:</strong> The harness records structured session transcripts (such as JSONL traces) capturing the exact prompt payload, the model’s internal thinking traces, the precise tool arguments sent, and the raw stdout/stderr returned by the sandbox.</li><li><strong>Context &amp; Token Telemetry:</strong> Real-time instrumentation tracking how context size inflates turn by turn. This lets you pinpoint runaway token spikes, prompt cache hit rates, and the exact cost of each iteration before a loop drains your budget.</li><li><strong>Enterprise OpenTelemetry &amp; Cloud Trace:</strong> When moving from local IDE harnesses to cloud-hosted agents (e.g. running Google ADK agents on Cloud Run or the Google Agent Platform), the harness exports distributed traces into <strong>Google Cloud Trace</strong> and structured events into <strong>Cloud Logging</strong>. Every tool call becomes an inspectable trace span.</li></ol><p>When you can inspect the full trajectory, debugging ceases to be guesswork. You can conduct a forensic post-mortem on any failed run and immediately determine: <em>Did the model experience a reasoning failure, or did my harness fail to provide the right deterministic rail?</em></p><h3>Summary &amp; What’s Next: Firing Up the Engine</h3><p>We have <em>shifted left</em> and engineered the deterministic foundations: living rules, version-controlled state, curated tools, progressive skill disclosure, plugin packaging, sandboxed boundaries, and execution traces. We have built the race car chassis.</p><p><em>Now, how does work actually get done?</em></p><p>In <strong>Part 2</strong>, we fire up the dynamic runtime engine inside the harness: <strong>Loop Engineering</strong>. We will explore:</p><ul><li>Why autonomous feedback loops crush single-pass generation.</li><li>The core mechanical components of an engineered developer loop.</li><li>The Ratchet Principle.</li><li>The Diagnostic Rule of Thumb: <em>Is it a Loop bug or a Harness bug?</em></li><li>Designing Evaluation Rubrics (combining traditional tests with LLM goal evaluation and deterministic tool trajectory verification).</li><li>A complete, turn-by-turn trace of an autonomous agent self-correcting in real time.</li><li>The loopy skill, and the Loop Library.</li></ul><p>👉 <strong>Continue to Part 2: Loop Engineering in the Harness: Autonomous, Self-Correcting Systems (Coming Soon)</strong></p><p>See you there!</p><h3>Before You Go</h3><ul><li><strong>Please share</strong> 📢 this with anyone that you think will be interested. It might help them, and it really helps me!</li><li>Please <strong>give me loads of claps</strong>! 👏 (Just hold down the clap button.)</li><li>Please <strong>leave a comment</strong> 💬. Interaction is good!</li><li><strong>Add a star</strong> ⭐ on my repos!</li><li><strong>Follow</strong> 👉 and <strong>subscribe</strong> 🔔, so you don’t miss my content.</li></ul><h3>References and Useful Links</h3><h4>The Harness &amp; Ecosystem</h4><ul><li><a href="https://github.com/derailed-dash/harness-engineering-demo"><em>Dazbo’s Harness Engineering Demo Application: no harness vs harness engineering, side-by-side</em> | GitHub</a></li><li><a href="https://cloud.google.com/discover/agent-harness?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco"><em>What is an agent harness? </em>| Google Cloud Discover</a></li><li><a href="https://www.youtube.com/watch?v=W9BX0jyzd2k"><em>What is an Agentic Harness?</em> (Google Cloud Tech)</a></li><li><a href="https://www.youtube.com/watch?v=F8EZJAm9iO8"><em>Harness Engineering Explained: Inside the Stack Behind Antigravity, Claude Code &amp; Cursor</em> (The Agent Factory with Ryan Lopopolo on Google Cloud Tech)</a></li><li><a href="https://medium.com/google-cloud/the-ai-native-developer-experience-part-2-harness-engineering-with-google-antigravity-7fb72dab243f"><em>The AI-Native Developer Experience, Part 2: Harness Engineering with Google Antigravity</em> by Jaroslav Pantsjoha (JP)</a></li><li><a href="https://www.builder.io/blog/de-slop-ai-generated-codebase"><em>How to De-Slop an AI-Generated Codebase</em> by Alice Moore | Builder.io</a></li><li><a href="https://github.com/google/skills/tree/main/plugins/cloud/google-cloud-developer">google-cloud-developer plugin on GitHub</a></li><li><a href="https://github.com/google/agents-cli"><em>Google Agents CLI </em>on GitHub</a></li><li><a href="https://github.com/vercel-labs/skills"><em>Vercel Skills CLI </em>on GitHub</a></li></ul><h4>Dazbo’s Related Articles</h4><ul><li><a href="https://medium.com/google-cloud/loop-engineering-in-the-harness-autonomous-self-correcting-systems-95d3ac8f64cf?postPublishedType=repub"><em>Part 2: Loop Engineering in the Harness: Autonomous, Self-Correcting Systems</em></a></li><li><a href="https://dev.to/gde/skills-sprawl-when-too-much-of-a-good-thing-confuses-your-ai-agent-4nij"><em>Skills Sprawl: When Too Much of a Good Thing Confuses Your AI Agent</em> by Dazbo</a></li><li><a href="https://dev.to/gde/dialling-our-agents-to-11-agent-skills-you-need-to-be-using-2o2k"><em>Dialling Our Agents to 11: Agent Skills You Need to be Using!</em> by Dazbo</a></li><li><a href="https://dev.to/gde/dialling-our-agents-to-11-my-favourite-mcp-servers-3hbm"><em>Dialling Our Agents to 11: My Favourite MCP Servers</em> by Dazbo</a></li><li><a href="https://dazbo.co.uk"><em>Dazbo’s Portfolio</em></a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6e7e0443a4fc" width="1" height="1" alt=""><hr><p><a href="https://medium.com/google-cloud/from-vibing-chaos-to-reliable-agents-the-what-why-and-how-of-harness-engineering-6e7e0443a4fc">From Vibing Chaos to Reliable Agents: The What, Why, and How of Harness Engineering</a> was originally published in <a href="https://medium.com/google-cloud">Google Cloud - Community</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Meet the Antigravity Extension: Gemini Code Assist on Steroids, in Whatever IDE You Love]]></title>
            <link>https://medium.com/google-cloud/meet-the-antigravity-extension-gemini-code-assist-on-steroids-33367e5a037c?source=rss-58a0bd8ba726------2</link>
            <guid isPermaLink="false">https://medium.com/p/33367e5a037c</guid>
            <category><![CDATA[antigravity-extension]]></category>
            <category><![CDATA[google-antigravity]]></category>
            <category><![CDATA[vscode]]></category>
            <dc:creator><![CDATA[Dazbo (Darren Lester)]]></dc:creator>
            <pubDate>Fri, 21 Aug 2026 01:11:12 GMT</pubDate>
            <atom:updated>2026-08-25T09:21:40.080Z</atom:updated>
            <content:encoded><![CDATA[<p>Google Antigravity (or Agy) was first launched as an agent-first development environment in late 2025. Then, in May 2026, Google released the new Antigravity suite, comprising four <em>surfaces</em>:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4w8OYWpC8QII9ANhctA0qQ.png" /><figcaption>Antigravity surfaces</figcaption></figure><ol><li><strong>Antigravity 2.0</strong> — The agent-first, intent-driven surface. For those who care about the outcomes, but not necessarily the code. It’s the modern agentic environment for <em>builders</em>.</li><li><strong>Antigravity IDE</strong> — The familiar IDE (the IDE itself is a fork of VS Code), but with all the smarts of the Antigravity harness built-in.</li><li><strong>Antigravity CLI</strong> — The terminal user interface (TUI) experience. This replaced the very awesome <strong>Gemini CLI</strong>.</li><li><strong>Antigravity SDK</strong> — For developers who want to harness the full power of Antigravity programmatically.</li></ol><p>All these surfaces use the same Antigravity <em>harness</em>. They give you capabilities such as:</p><ul><li>Advanced multi-stage reasoning, using the latest Gemini models.</li><li>Full repo refactoring.</li><li>Planning, execution and evidencing.</li><li>Agent loops and subagent swarm orchestration.</li><li>Integration with file system and shell environment.</li><li>Multiple useful built-in commands and subagents.</li></ul><p>At the same time, Google announced that both <a href="https://docs.cloud.google.com/gemini/docs/codeassist/overview?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Gemini Code Assist</a> and <a href="https://geminicli.com/">Gemini CLI</a> were <a href="https://developers.googleblog.com/an-important-update-transitioning-gemini-cli-to-antigravity-cli/?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco#important-timeline-for-consumers">end-of-life</a>. And indeed, unless you happened to run inside an organisation that had paid for enterprise licenses, these products got turned off in June 2026.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*E-9aVQRkfnhKVOEw8OBswg.png" /><figcaption>Gemini CLI — end-of-life</figcaption></figure><h3>The Worry</h3><p>Gemini CLI was great. I was gutted to see it go. But <a href="https://antigravity.google/product/antigravity-cli?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Antigravity CLI</a> was a worthy replacement, and it was immediately available. So we made the switch, and we were happy.</p><p><strong>But what about the death of Gemini Code Assist?</strong> If you’re not familiar with this tool: it was the Gemini agent that you could install as an extension into your favourite IDE, like VS Code or IntelliJ. The agent ran in a side panel, fully-integrated with your IDE.</p><p>Then Google told us:</p><p><em>“Gemini Code Assist is dead. But don’t worry. For all you coders out there, just switch to Antigravity IDE. It feels like VS Code and you get the Gemini Code Assist experience — on steroids.”</em></p><p>(They may not have actually used these words.)</p><p>And this is exactly what I did. I briefly mourned VS Code, and then went all-in on Agy IDE.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7VEbJ5ibRsnsb82lBGCZJg.png" /><figcaption>All in on Agy IDE</figcaption></figure><p>But I was rather vocal in sharing my thoughts with some influential folk at Google.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_1NrgA5Q2ka1giBUweh5Zg.png" /><figcaption>Sharing opinions</figcaption></figure><p>My thoughts being:</p><ol><li>Some folks — like me — were happy to make the switch. But there are tens of millions of developers who <strong>WILL NOT give up their favourite IDE</strong>. So, if you tell them: <em>“Switch to Agy, or you stay where you are, but with no Gemini”</em>, then they’re gonna say: <em>“Okay. I’ll go with option 2.”</em> To me, that’s a lose-lose situation.</li><li>Some enterprises are pretty strict about their approved development environments and ecosystems. And — for reasons — they can be SOOOOO SLOW at adopting new things.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aQJdsaarR5nTs1eFtbdg7Q.png" /><figcaption>So slow</figcaption></figure><p>The TL;DR — there was no viable path for developers who wanted to evolve from the Gemini Code Assist experience, but stay with their favourite IDE.</p><h3>Until Now</h3><p>Today (August 20, as I write this), Google launched the Antigravity extension for IDEs! You can check out the official Google release blog, <a href="https://antigravity.google/blog/antigravity-ide-extensions?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">here</a>. Right now, there are Agy extensions for VS Code, Visual Studio, IntelliJ IDEs, and Zed.</p><p><strong>I can’t tell you how stoked I am about this announcement!</strong></p><p>What is it? Well, think of it as a replacement for the Gemini Code Assist extension. It’s Gemini Code Assist, but pumped with steroids and caffeine. It’s all the cool Agy harness capabilities we’ve come to love in Antigravity 2.0, Antigravity CLI and Antigravity IDE, but now packaged as an extension.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Mxzgyjb4kCL-QJXEKvxzfg.png" /><figcaption>Antigravity extension</figcaption></figure><p>So, as of tonight, I’ve switched back to VS Code!</p><h3>Install the Extension</h3><p>Let’s install it! I’m using VS Code, so I’m installing <a href="https://marketplace.visualstudio.com/items?itemName=Google.google-antigravity">this extension</a>.</p><p>If you happen to go looking for “Google Antigravity” in the VS Code Marketplace, be sure to install this one by google.com:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EgGvO2mfitvjslm-I9tX1A.png" /><figcaption>Installing Antigravity extension</figcaption></figure><p>When you open the extension, you’ll see this in VS Code:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bnwherE26RXoZA5NPU4jEg.png" /><figcaption>Install in VS Code</figcaption></figure><p>Click <strong>Install</strong>, and you’ll be asked to sign-in to Google:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oadIeaoSuNzZ8uPBaPs9dA.png" /><figcaption>Sign-in</figcaption></figure><p>And you’ll be asked if you want to install any of the ready-made <strong>plugins</strong>:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*D49GEKvtoIW7ypvogSDEZA.png" /><figcaption>Agy plugins</figcaption></figure><p>And that’s it!</p><p>It took about 20 seconds, and now the Antigravity extension is working in my VS Code!</p><h3>Take It For a Spin</h3><p>I wondered if all my existing agent skills and MCP servers would be working. So I asked it: <em>“Install my Gemini code reviewer”</em></p><p>It immediately found and activated my <a href="https://github.com/derailed-dash/dazbo-agent-skills">install-gemini-code-review-action skill</a> and used it to install my <a href="https://github.com/derailed-dash/gemini-review-action">Gemini Code Review Action</a> into the local repo. It then asked for permission to commit these changes push them to GitHub. I gave Agy permission, and it went ahead and pushed.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/745/1*pVMups2VS5VRYS48OAQYQQ.png" /><figcaption>Activated skills and MCPs</figcaption></figure><p>So this proved that my existing skills were immediately discovered, activated and used, and that it was able to use my GitHub MCP to perform the commit and push. Nice!</p><h3>Models</h3><p>We can see the models that are available:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/813/1*o0j8NqvhZYD_nZc-vGMlpw.png" /><figcaption>Model selection in Agy</figcaption></figure><p>And we can see more model and Antigravity extension <strong>settings </strong>here:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ywt2a2N7HNp9ByLjYDLvDQ.png" /><figcaption>Settings</figcaption></figure><p><em>(I love the fact that next to my “AI Ultra” subscription is an “Upgrade” button. Seriously? Lol!)</em></p><h3>The Plugins Created Some Skills Duplication</h3><p>The newly installed plugins resulted in duplication of some skills I had previously installed in my ~/.gemini/config/skills folder. So I said to Antigravity: <em>&quot;Let&#39;s do a skills optimisation run.&quot;</em></p><p>It went ahead and activated my <a href="https://github.com/derailed-dash/dazbo-agent-skills">organise-agent-skills</a>. The skill made a number of recommendations which I implemented. The results looked like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/835/1*d49mAXiDZpOtplaXO1JWhQ.png" /><figcaption>Skills optimised</figcaption></figure><p>Wow!</p><ul><li>29 duplicated skills deleted.</li><li>56 “sub-skills” set to excluded, because they are managed by their parent skills.</li><li>68% reduction in token overhead!</li></ul><p>Great, so our skills are back to running nice and lean.</p><h3>A Nice View of Token Usage</h3><p>Whilst we’re talking about token optimisation, I stumbled across this “Customization” page in the Agy extension settings:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uAhSP9KXVmgTtns38hEVDg.png" /><figcaption>Agy Settings — Customizations</figcaption></figure><p>This is nice view that how much of your “customization budget” is being consumed by all your GEMINI.md (or AGENTS.md), your skills, and your MCP tools. What’s quite interesting to note is that Agy will automatically start truncating if you exceed this budget. So you might suddenly find that your skills are being ignored, for example. This is another good reason to optimise your install skills and MCP servers.</p><h3>Things Not Working</h3><h4>Browser Subagent</h4><p>The first thing I struggled with is the built-in browser subagent. I couldn’t get this to browse web pages for me, even after I enabled the remote debugging in Chrome. I did expect this. My VS Code is running in Windows, but connected to my repos in WSL. This has typically been a problem area for browser subagents. Maybe I’ll be able to get it working, if I put in a bit of work. But tonight… I’m far too tired!</p><p><em>(I should mention: this issue is only for WSL users. 4 out of 5 developers won’t be affected by this.)</em></p><p>On the plus side, I was able to remote control the browser with my existing <a href="https://browsermcp.io/">BrowserMCP</a> configuration. No changes required.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/868/1*M3ofDBWt6cYCwC7hmWT5UA.png" /><figcaption>Browsing with BrowserMCP extension</figcaption></figure><h4>Antigravity In-Line Complete and Tab Features</h4><p>One thing I really love about Antigravity IDE is the real-time suggestions and edits you can make in the editor window.</p><p>For example: if I were to change one instance of <em>“We did something”</em> to <em>“I did something”</em>, then Antigravity IDE allows me to update all remaining instances in a context-aware way, just by pressing tab.</p><p>This feature is not currently working for me with the Agy extension in VS Code. So it’s possible my configuration isn’t correct, or this feature hasn’t been implemented in the extension yet. But this is only day 1!</p><h3>What Does This Mean for Agy IDE?</h3><p>Within hours of the extension arriving, many folks have been asking me, <em>“What does this mean for Agy IDE? Will Google kill it?” </em>It’s a question I’ve been asking my friends at Google for a little while.</p><p><em>Nobody </em>has explicitly told me that Google <em>will </em>kill Agy IDE. But my thoughts are:</p><ol><li>If the Agy extension provides the full Agy harness experience in all the major IDEs, what’s the <em>point </em>in having a separate IDE that looks a lot like VS Code?</li><li>Why would Google want to maintain the “IDE” components of Agy IDE, when these are essentially just a fork of VS Code anyway? That’s a lot of work.</li><li>Lastly, take a look at the <a href="https://antigravity.google/changelog">release notes</a> for Agy 2.0, Agy CLI, and Agy IDE. The relative sparsity (is that a word?) of Agy IDE patches is telling.</li></ol><p>Personally, I think the writing is on the wall.</p><h3>Wrap-Up</h3><p>I’m so happy this Agy extension has arrived. Now any developer can use the extremely capable Antigravity harness from their IDE of choice. It gives a viable migration path for those who were using Gemini Code Assist before.</p><p>For me, the install experience was totally painless, and everything just worked out-of-the-box. (Well, almost everything. But I can wait for the other things.) The agent is responsive and works. And it picked up all my existing configuration.</p><p>That’s it! Go forth and use Antigravity in your favourite editor!</p><h3>Before You Go</h3><ul><li><strong>Please share</strong> 📢 this with anyone that you think will be interested. It might help them, and it really helps me!</li><li>Please <strong>give me loads of claps</strong>! 👏 (Just hold down the clap button.)</li><li>Please <strong>leave a comment</strong> 💬. Interaction is good!</li><li><strong>Add a star</strong> ⭐ on my repos!</li><li><strong>Follow</strong> 👉 and <strong>subscribe</strong> 🔔, so you don’t miss my content.</li></ul><h3>References and Useful Links</h3><ul><li><a href="https://dazbo.co.uk/">Dazbo’s Portfolio</a></li><li><a href="https://antigravity.google/?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Google Antigravity Official Site</a></li><li><a href="https://antigravity.google/blog/antigravity-ide-extensions?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Google Antigravity IDE Extensions Announcement</a></li><li><a href="https://marketplace.visualstudio.com/items?itemName=Google.google-antigravity">Google Antigravity Extension on the Visual Studio Marketplace</a></li><li><a href="https://github.com/derailed-dash/dazbo-agent-skills">Dazbo’s Agent Skills Repository</a> — if you like it, please add a star!</li><li><a href="https://github.com/derailed-dash/gemini-review-action">Dazbo’s Gemini Code Review Action</a> — if you like it, please add a star!</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=33367e5a037c" width="1" height="1" alt=""><hr><p><a href="https://medium.com/google-cloud/meet-the-antigravity-extension-gemini-code-assist-on-steroids-33367e5a037c">Meet the Antigravity Extension: Gemini Code Assist on Steroids, in Whatever IDE You Love</a> was originally published in <a href="https://medium.com/google-cloud">Google Cloud - Community</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Skills Sprawl: When Too Much of a Good Thing Confuses  Your AI Agent]]></title>
            <link>https://medium.com/google-cloud/skills-sprawl-when-too-much-of-a-good-thing-bloats-your-ai-agent-167f907e9dc2?source=rss-58a0bd8ba726------2</link>
            <guid isPermaLink="false">https://medium.com/p/167f907e9dc2</guid>
            <category><![CDATA[agent-skills]]></category>
            <category><![CDATA[progressive-disclosure]]></category>
            <category><![CDATA[model-quota]]></category>
            <category><![CDATA[google-antigravity]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <dc:creator><![CDATA[Dazbo (Darren Lester)]]></dc:creator>
            <pubDate>Mon, 17 Aug 2026 02:21:46 GMT</pubDate>
            <atom:updated>2026-08-17T07:06:38.849Z</atom:updated>
            <content:encoded><![CDATA[<h3>Skills Sprawl: When Too Much of a Good Thing Confuses Your AI Agent</h3><h3>Introduction</h3><p>In this post, we’re going to look at the problem of <strong>Skills Sprawl</strong>.</p><p><em>What happens if you have too many skills?</em></p><p>We’ll understand LLM decision fatigue and tool selection accuracy, inspect the mechanics of excluded vs inactive skills in Google Antigravity, and see how pruning your setup can make your agent smarter and save hundreds of thousands of tokens per session.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*A76X2sJgjDw-FjB3GtYo3g.png" /><figcaption>I have a special set of skills</figcaption></figure><h3>Skills Context (It’s a Sort of Pun…)</h3><p>Everyone knows that <em>agent skills</em> are awesome. If you don’t, then:</p><ol><li>Where have you been?</li><li>You’re in for a treat!</li></ol><p>In some of my previous posts I’ve talked about how skills act as <strong>on-demand power-ups</strong> for our agents. We use them to provide knowledge, rules, and workflows to do things the agent (or more specifically, the model) didn’t otherwise know how to do effectively.</p><p>You might have seen me compare this to how Neo loads his skills in the Matrix. <em>“I know Kung Fu!”</em></p><p>This <em>just-in-time knowledge</em> provides a number of advantages to our agents:</p><ul><li>They know how to do a thing well.</li><li>They don’t hallucinate the things they don’t know about.</li><li>They don’t have to experiment and course-correct; consequently, they’re more likely to succeed the first time, and they do it with a minimum number of tokens consumed.</li></ul><p>In short: <strong>faster, cheaper, and more reliable</strong>.</p><p>Quick aside: when I say <em>agent</em>, I’m typically talking about your agentic partner in development crime, like <a href="https://antigravity.google/?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Google Antigravity</a>, Claude Code, or whatever. But when you’re building your own agentic solutions, you can use skills in exactly the same way. For example, check out my blog <a href="https://medium.com/google-cloud/automated-github-code-reviewsusing-google-gemini-7b4d027b3092?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Automated GitHub Code Reviews Using Google Gemini</a>, where I’ve built an agentic PR review solution that leverages skills.</p><h3>Overview of Progressive Disclosure</h3><p>Skills use a cool mechanism called <strong>progressive disclosure</strong> to load on-demand.</p><p>Let’s quickly recap this mechanism:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yczy0Bc7EMcdlt5ynA8sNQ.png" /><figcaption>Progressive Disclosure</figcaption></figure><h4>Level 1: Metadata (Frontmatter)</h4><p>When your agent starts up, it reads <strong>the frontmatter</strong> — and <strong>only the frontmatter</strong> — of <strong>every installed skill</strong> in your global configuration and workspace.</p><p><em>By the way, if you don’t know where these skills should be placed, or you don’t know the difference between global and workspace locations, check out my earlier blog: </em><a href="https://medium.com/google-cloud/confused-about-where-to-put-your-agent-skills-ea778f3c64f3?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco"><em>Confused About Where to Put Your Agent Skills?</em></a></p><p>The frontmatter is simply the descriptive YAML at the top of each skill which contains the skill name, a description (what it does and when it should be used), and any additional useful metadata (such as what tools it might use, author, etc).</p><p>For example, this is from my maintaining-core-documentation skill:</p><pre>name: maintaining-core-documentation<br>description: |<br>  Creates, maintains, and synchronises core project documentation <br>  (README, TODO, DESIGN, Architecture, Testing, Deployment). <br>  Use when the user needs to write, update, or structure project documentation <br>  based on codebase changes, or whenever asked to perform documentation <br>  reviews or updates.<br>metadata:<br>  author: Darren &quot;Dazbo&quot; Lester<br>  repository: https://github.com/derailed-dash/dazbo-agent-skills</pre><p>Crucially, this frontmatter section is <em>tiny</em>. So even though the agent reads it up-front, it consumes very few tokens. Typically fewer than 100 tokens per skill.</p><p>The agent now has a mental model of all the skills that are available to it.</p><h4>Level 2: Instructions Loading (On-Demand)</h4><p>When the agent is asked to do a task that one of these skills might be useful for, the model selects the skill (based on the frontmatter), and then <strong>activates</strong> it. When this happens, the agent reads the SKILL.md body into context.</p><p>This will typically add several hundred, if not thousands of tokens.</p><h4>Level 3: Resources and Code (As Needed)</h4><p>Some skills may package additional resources, such as scripts, reference documents, or other assets. For example, my maintaining-core-documentation skill has a references subdirectory with several document template files.</p><p>These assets are only loaded as needed.</p><h4>How This Helps</h4><p>When we have skills, we don’t need to include massive quantities of instructions in our global GEMINI.md or AGENTS.md files. Instead, all this knowledge is packaged into individual, self-contained, portable, on-demand units of knowledge.</p><p>It is an elegant, modular pattern that massively reduces your token consumption.</p><h3>But Can You Have Too Much of a Good Thing?</h3><p>There’s no shortage of skills out there! You can use a public hub like <a href="https://skills.sh/">skills.sh</a> to find just about any skill you need. And, of course, you can create your own.</p><p>Check out my blog <a href="https://dev.to/gde/dialling-our-agents-to-11-agent-skills-you-need-to-be-using-2o2k"><em>Dialling Our Agents to 11: Agent Skills You Need to be Using!</em></a>, where I list some of my favourite skills.</p><p>The problem is… You end up <em>collecting.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EUSL_bT3KUkKu23kbVSiIw.png" /><figcaption>Collecting Skills</figcaption></figure><p>I recently asked Antigravity to tell me how many skills I have. (I have sarcasm dialled up…)</p><pre>Well, Dazbo, according to my context window — which is currently groaning <br>under the crushing weight of your digital hoard — you currently have a <br>staggering 170 skills installed.<br><br>Because obviously, why settle for a sleek, efficient setup when you can <br>force your AI assistant to lug around the instruction manual for every <br>obscure file converter, niche API, and hypothetical cloud deployment <br>scenario known to humanity?<br><br>Truly, nothing says &#39;peak productivity&#39; quite like burning thousands of <br>tokens before we&#39;ve even started speaking, just so I know how to format a <br>Dev.to post while simultaneously standing by to orchestrate a multi-region <br>Spanner failover.<br><br>I&#39;m sure you&#39;ll be adding a 171st skill any second now - presumably <br>one to help you manage the utter chaos of the first 170.</pre><p>Hilarious. 170. <em>Maybe I have too many.</em></p><p>The irony here is that Antigravity’s last comment was spot on. I decided to <strong>build a skill to organise and optimise my skills!</strong></p><h3>What’s the Problem?</h3><p>As we’ve already covered: with progressive disclosure, the agent framework performs <strong>Tier 1 Discovery</strong> on startup. It parses the frontmatter of <strong>every installed skill</strong> and injects this straight into the system prompt.</p><p>A well-crafted skill description needs enough context, trigger keywords, and intent markers for the LLM to know when to select it. (<em>“Use when…”</em>)</p><p>When I analysed my 170 skills, it turns out that my average skill frontmatter contained 103 tokens. Now let’s do some quick maths:</p><p>170 active skills × 100 tokens/description =17,000 tokens</p><p>That means my agent is injecting <strong>17,000 tokens of system prompt overhead on every single turn</strong>.</p><p>If I’m engaged in a typical 30-turn pair-programming session:</p><p>30 turns × 17,000 tokens = 510,000 tokens</p><p>That means I am burning an additional <strong>half a million tokens</strong> just to re-read the index of my skill library over and over again.</p><p><em>So what?</em></p><p>Let’s look at some of the potential issues this creates:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jrcohdz2EZFyc4EgQTYJHQ.png" /><figcaption>Issues of skills sprawl</figcaption></figure><ul><li>Depending on your model choice, caching configuration, and AI subscription, there may be <strong>financial implications</strong>. But it’s likely to be small.</li><li>It will definitely <strong>eat into your token limits</strong>. Most models and plans have token limits and rate limits. Adding an extra 17K tokens per turn will definitely reduce how long it takes for you to exhaust your quota!</li><li>Most importantly, <strong>it’s confusing your model</strong>.</li></ul><h3>Confusion-Nation</h3><p>Skills sprawl causes <strong>decision fatigue and skill selection accuracy degradation</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rS0wAE_uhallocyBpMfZXw.png" /><figcaption>Skill selection confusion</figcaption></figure><p>When a model is presented with 170 potential skills simultaneously, it faces <strong>choice ambiguity</strong>. Many skills will naturally have overlapping domain descriptions. For example, my 170 skills had overlaps in each of these areas:</p><ul><li>Content generation</li><li>API guides (particularly for Gemini APIs and SDKs)</li><li>Google products and services</li><li>Test-driven development (TDD) and code reviews</li></ul><p>Does academic research back this up? Absolutely.</p><h4>The Empirical Evidence: Shortlist Depth and Selection Accuracy</h4><p>In a recent paper titled <a href="https://arxiv.org/pdf/2605.24660"><em>“How Many Tools Should an LLM Agent See? A Chance-Corrected Answer”</em></a> (<em>Repantis et al</em>, May 2026), researchers evaluated LLM tool selection performance across 370 tools using the <a href="https://gorilla.cs.berkeley.edu/leaderboard.html"><strong>Berkeley Function Calling Leaderboard (BFCL)</strong></a> — the industry-standard benchmark for evaluating model tool-calling capabilities.</p><p>They concluded:</p><p><em>“Show too many tools and the model struggles to choose. Show too few and the correct tool may not appear.”</em></p><p>There is an important nuance here: if you only give a model 2 choices, it has a 50% chance of guessing correctly by pure luck. If you show it 50 choices, blind luck drops to 2%.</p><p>When the researchers <strong>corrected for random chance</strong> — measuring genuine model comprehension rather than lucky guesses — the impact of list size was striking:</p><ul><li>Presenting an LLM with a crowded shortlist of <strong>50 tools</strong> resulted in <strong>60.9% selection accuracy</strong>.</li><li>Scoping that shortlist down to <strong>~7 focused tools</strong> boosted true accuracy to <strong>76.8%</strong> (a 15.9% leap in decision precision).</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eKmze41wzoMkaIIHs1vNXw.png" /><figcaption>Decision-making accuracy</figcaption></figure><p>Related benchmarks like <a href="https://arxiv.org/abs/2310.03128"><strong>MetaTool</strong></a> and <a href="https://arxiv.org/abs/2307.16789"><strong>ToolBench</strong></a> show the same results: as toolset sizes scale past 30–50 items without structured routing, error rates spike rapidly due to <strong>description collision</strong> and noise. The model gets confused by keyword overlaps and either picks the wrong tool or hallucinates non-existent parameters.</p><p>Okay, in this blog, I’m talking about skills, not tools. But the principle is the same. Too many skills lead to the same problems of confusion caused by description collision.</p><p>We want:</p><ul><li>Our agent to pick the right skills with a high degree of accuracy.</li><li>To avoid burning unnecessary tokens.</li></ul><p><strong>I need a way to optimise my installed skills!</strong></p><h3>Why Do I Have So Many Skills?</h3><p>Before I start pruning my skills, I need to understand why I have so many.</p><p>I went back and reviewed my Git history to trace how I got here. Here’s what I discovered:</p><ol><li><strong>Google Cloud &amp; AI knowledge:</strong> Over 120 skills pulled from official Google repositories, including <a href="https://github.com/google/skills">google/skills</a>, <a href="https://github.com/google/agents-cli">google/agents-cli</a>, <a href="https://github.com/google-gemini/gemini-skills">google-gemini/gemini-skills</a>, and <a href="https://github.com/GoogleCloudPlatform/vertex-ai-creative-studio/tree/main/experiments/mcp-genmedia/skills">GoogleCloudPlatform/vertex-ai-creative-studio</a>. These include Google products and services (like BigQuery, Cloud SQL, GKE, Cloud Run), the Google Cloud Well-Architected Framework (WAF), Gemini and Google GenAI APIs and SDKs, and GenMedia.</li><li><strong>Dazbo Agent Skills:</strong> Custom skills I wrote for <a href="https://github.com/derailed-dash/dazbo-agent-skills">derailed-dash/dazbo-agent-skills</a> for documentation, blogging, security, installing automated PR code reviews, skill organisation, and deployment.</li><li><strong>Core Software Engineering &amp; Best Practices:</strong> 25 skills adopted from Addy Osmani’s <a href="https://github.com/addyosmani/agent-skills">addyosmani/agent-skills</a> repository covering TDD, code review, debugging, interface design, context engineering, and software delivery workflows.</li><li><strong>Research, Strategy &amp; Technical Writing:</strong> 5 specialised skills installed from Shubham Saboo’s <a href="https://github.com/Shubhamsaboo/awesome-llm-apps/tree/main/agent_skills">shubhamsaboo/awesome-llm-apps</a> repository (deep-research, fact-checker, strategy-advisor, technical-writer, content-creator).</li><li><strong>Specialised &amp; Community Skills:</strong> Niche skills sourced from specific community repositories, including <a href="https://github.com/wshobson/agents">wshobson/agents</a> (documentation-and-adrs, interview-me), <a href="https://github.com/coreyhaines31/marketingskills">coreyhaines31/marketingskills</a> (seo-audit), <a href="https://github.com/remotion-dev/skills">remotion-dev/skills</a> (remotion-best-practices), and <a href="https://github.com/vercel-labs/skills">vercel-labs/skills</a> (find-skills).</li></ol><p>Here is the complete breakdown of the 170 skills I had installed, grouped by category and origin repository:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fsiteproxy.ruqli.workers.dev%3A443%2Fhttps%2Fdatawrapper.dwcdn.net%2FK4Fu7%2F4%2F&amp;display_name=Datawrapper&amp;url=https%3A%2F%2Fsiteproxy.ruqli.workers.dev%3A443%2Fhttps%2Fdatawrapper.dwcdn.net%2FK4Fu7%2F4%2F&amp;image=https%3A%2F%2Fsiteproxy.ruqli.workers.dev%3A443%2Fhttps%2Fdatawrapper.dwcdn.net%2FK4Fu7%2Fplain-s.png%3Fv%3D4&amp;type=text%2Fhtml&amp;schema=dwcdn" width="600" height="1853" frameborder="0" scrolling="no"><a href="https://medium.com/media/7eabf683ff134a7fff26c16d83eecbc0/href">https://medium.com/media/7eabf683ff134a7fff26c16d83eecbc0/href</a></iframe><p>So you can see how easy it is for your skills base to get out of hand! Especially if you work with Google Cloud and Google AI services like I do. It’s super easy to install over 100 Google-related skills, by just following a few Google blog recommendations.</p><h3>Next Question: Do I Need Them All?</h3><p>I asked Antigravity to read all of my skills in detail, and identify any areas of overlap and redundancy.</p><p>There was a LOT of redundancy! We uncovered seven major categories of overlap, duplication, and inefficiency:</p><ol><li><strong>Exact Byte-for-Byte File &amp; Naming Duplicates:</strong> In <a href="https://github.com/google/skills/tree/main/skills/cloud/gemini-agents-api">google/skills</a>, the skill directory gemini-agents-api declared name: gemini-managed-agents-api in its frontmatter. This directory vs. frontmatter mismatch caused installation tools to spawn two exact duplicate folders (gemini-agents-api and gemini-managed-agents-api) carrying 100% byte-for-byte identical content.</li><li><strong>Deprecated SDKs vs. Modern APIs:</strong> Older skills still provided code snippets using the deprecated google-generativeai package and obsolete model strings, directly contradicting modern google-genai SDK standards and confusing the agent during code generation.</li><li><strong>API Guide Overlaps &amp; Fragmentation Across Repositories:</strong> The gemini-api skill in <a href="https://github.com/google/skills/tree/main/skills/cloud/gemini-api">google/skills</a> (which exists as an enterprise Agent Platform guide) collided directly with gemini-api-dev in <a href="https://github.com/google-gemini/gemini-skills/tree/main/skills/gemini-api-dev">google-gemini/gemini-skills</a>. Installing across multiple Google repositories resulted in competing guides for the same underlying APIs (gemini-api, gemini-api-dev, gemini-agents-api, gemini-interactions-api, and gemini-live-api-dev).</li><li><strong>Irrelevant Domain Bundles (The Monolithic Repo Problem):</strong> Pulling down the monolithic <a href="https://github.com/google/skills">google/skills</a> repository imported mobile advertising suites (Android/iOS banner ads, Unity SDKs) that had zero relevance to cloud backend architecture and agent development.</li><li><strong>Hierarchical Sub-Skill Redundancies (The Meta-Skill Tax):</strong> 25 individual SDLC skills from <a href="https://github.com/addyosmani/agent-skills">addyosmani/agent-skills</a> (such as test-driven-development and code-review-and-quality) were loaded into system prompt context at startup, despite being child sub-skills already orchestrated on-demand by using-agent-skills. Similarly, 20 sub-skills from <a href="https://github.com/google/agents-cli">google/agents-cli</a> were loaded continuously rather than routed via google-agents-cli-workflow.</li><li><strong>Tooling &amp; MCP Surface Overlaps:</strong> Standalone tools like the adk-docs-mcp server were redundant because google-developer-knowledge already indexes ADK documentation (adk.dev).</li><li><strong>Capability &amp; Persona Redundancies:</strong> documentation-and-adrs (<a href="https://github.com/addyosmani/agent-skills">addyosmani/agent-skills</a>) was a duplicate of architecture-decision-records (<a href="https://github.com/wshobson/agents">wshobson/agents</a>), while generic content-creator (<a href="https://github.com/Shubhamsaboo/awesome-llm-apps/tree/main/awesome_agent_skills">shubhamsaboo/awesome-llm-apps</a>) collided with my custom <a href="https://github.com/derailed-dash/dazbo-agent-skills">dazbo-content</a> persona.</li></ol><h3>What Else Did I Discover?</h3><p>Beyond the sheer volume of redundant skills, I made another cool discovery: <strong>many skill collections come with a “parent” Meta-Skill.</strong></p><p>When you install a large collection of skills — such as Addy Osmani’s engineering skills (<a href="https://github.com/addyosmani/agent-skills">addyosmani/agent-skills</a>), or Google&#39;s 20 Agent Platform skills (<a href="https://github.com/google/agents-cli">google/agents-cli</a>) — your instinct is to leave all 20 to 25 skills enabled in your workspace. I mean... <em>Why wouldn&#39;t you?</em></p><p>However, these skill suites are designed to be <strong>hierarchical</strong>:</p><ul><li><strong>addyosmani/agent-skills:</strong> Orchestrated by the using-agent-skills meta-skill.</li><li><strong>Google ADK &amp; Agent Platform:</strong> Orchestrated by the google-agents-cli-workflow meta-skill.</li><li><strong>GCP Data Pipelines:</strong> Orchestrated by the gcp-data-pipelines meta-skill.</li></ul><p>In each case, the parent meta-skill is supposed to work as a sort of skills index or decision tree. When a specific task comes in — such as writing unit tests or running a security audit — the parent meta-skill directs the agent to fetch and read <em>only</em> the specific <em>child</em> sub-skill from disk on demand.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wqIsg3X25VdwDI8XItDThA.png" /><figcaption>Parent meta-skill routing</figcaption></figure><p>Conclusion? <strong>We don’t need to load the frontmatter of all the <em>child</em> skills. We only need the parent skill!</strong> In the examples above, I can easily replace nearly 50 skills with just 3 orchestration skills!</p><p>So now I could implement a solution to organise my skills. The primary objectives:</p><ul><li>Remove skills that are truly redundant.</li><li>Prevent Level 1 skills loading (i.e. reading frontmatter into context) for skills that have parent meta-skills. For these, only pre-load the parent meta-skills.</li></ul><h3>Clarifying Skill States</h3><p>To design a lean agent workspace, we must be crystal clear on the three distinct states a skill can occupy throughout its lifecycle. They can be:</p><ul><li>Installed and inactive (discoverable).</li><li>Installed and excluded.</li><li>Activated.</li></ul><p>Let’s define these in more detail:</p><ol><li><strong>Installed and inactive (discoverable)</strong>: These skills are present on disk and discoverable by our agent. When the agent (e.g. Antigravity) starts, their frontmatter is automatically read into context.</li><li><strong>Installed and excluded</strong>: These skills reside in your disk library but are <em>excluded</em> from automatic Level 1 loading. For these, the agent does not automatically read their frontmatter and is therefore not <em>directly</em> aware of them during your conversation with it. But because they are present on disk and available to the agent, they can still be explicitly activated.</li><li><strong>Activated</strong>: This is the runtime state when a skill’s full SKILL.md body has been loaded into the current turn context, i.e. Level 2 loading (and Level 3, where supporting files are present and appropriate).</li></ol><p>Crucially, <strong>both <em>installed and inactive</em> and <em>installed and excluded</em> skills can transition into this <em>Activated</em> state.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jMH7ZpafWGnrP1SZRtsS1Q.png" /><figcaption>Skill states</figcaption></figure><h3>Excluding Skills</h3><p>So now we know that a good optimisation strategy is to use exclusion to prevent a bunch of skills being loaded at startup into the “Installed and inactive” state.</p><p>But how can we do this?</p><p>In <strong>Google Antigravity</strong>, global skill exclusions are managed via ~/.gemini/config/skills.json (or .agents/skills.json for workspace-level skills).</p><p>The native exclusion array accepts skill folder names. For example:</p><pre>{<br>  &quot;exclude&quot;: [<br>    &quot;alloydb-basics&quot;,<br>    &quot;cloud-spanner-migrations&quot;,<br>    &quot;firebase-basics&quot;<br>  ]<br>}</pre><p>Any skill listed in the exclude array is completely skipped during Tier 1 discovery. Its frontmatter is therefore <strong>not</strong> injected into the system prompt, saving tokens instantly.</p><p>So this is easy!</p><p>But here’s a cool trick. You can easily “un-exclude” a skill by prefixing its name with // inside the quotes. This is useful because it allows us to re-enable up-front skill loading, without having to check our disk to find the skill and its name.</p><p>So my approach is to always have <em>all</em> my available skills in the exclude list, but to comment-out the ones that need to be discovered automatically.</p><p>Note: I have the // <em>inside</em> the quotes. Why? Because this is not valid JSON:</p><pre>{<br>  &quot;exclude&quot;: [<br>    //&quot;alloydb-basics&quot;,<br>    //&quot;find-skills&quot;,<br>    &quot;firebase-basics&quot;<br>  ]<br>}</pre><p>But this is:</p><pre>{<br>  &quot;exclude&quot;: [<br>    &quot;//alloydb-basics&quot;,<br>    &quot;//find-skills&quot;,<br>    &quot;firebase-basics&quot;<br>  ]<br>}</pre><p>And because the exclude list uses <em>exact</em> string matching, you can use this // trick to enable/exclude skills on the fly, without having to remove the skill from the exclude list entirely.</p><p>In the example above:</p><ul><li>alloydb-basics and find-skills are <em>not</em> excluded, so they are auto-discovered on startup and their frontmatter is loaded into context.</li><li>firebase-basics <em>is</em> excluded, so it is <em>not</em> auto-discovered on startup and its frontmatter is <em>not</em> loaded into context.</li></ul><p>Now we’ve got everything we need to implement a repeatable mechanism to organise and clean up our skills.</p><h3>Irony and Guitars: I Made a Skill!</h3><p>Because: when you’ve got too many skills, you definitely need one more. (In this way, skills are a lot like guitars, or cats.)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*u9t-1Obl_yzzewHyQdbpNA.png" /><figcaption>Skills are like guitars and cats</figcaption></figure><p>Anyway, based on what I learned, I created a skill that:</p><ol><li>Identifies duplicative skills and recommends complete removal of those that are redundant.</li><li>Identifies which skills are orchestrated by a parent <em>meta-skill</em> and <em>excludes</em> the <em>child</em> skills from automatic discovery.</li><li>Identifies skills which look particularly niche or rarely used, and recommends excluding these also. They can always be included later, or even activated by explicitly asking the agent to do so.</li></ol><p>Let’s see how it performs! After installing my skill, I said to Antigravity:</p><p><em>“Optimise my skills please”</em></p><h3>Outcomes of Optimising</h3><p>My skill ran and presented this proposal summary table:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fsiteproxy.ruqli.workers.dev%3A443%2Fhttps%2Fdatawrapper.dwcdn.net%2FUmeh4%2F2%2F&amp;display_name=Datawrapper&amp;url=https%3A%2F%2Fsiteproxy.ruqli.workers.dev%3A443%2Fhttps%2Fdatawrapper.dwcdn.net%2FUmeh4%2F2%2F&amp;image=https%3A%2F%2Fsiteproxy.ruqli.workers.dev%3A443%2Fhttps%2Fdatawrapper.dwcdn.net%2FUmeh4%2Fplain-s.png%3Fv%3D2&amp;type=text%2Fhtml&amp;schema=dwcdn" width="600" height="367" frameborder="0" scrolling="no"><a href="https://medium.com/media/439280d0bfc5c8679c708113b715d30c/href">https://medium.com/media/439280d0bfc5c8679c708113b715d30c/href</a></iframe><p>It goes on to say:</p><pre>Upon your confirmation, I will execute the following steps:<br><br>1. **Retire Redundant Skills**: Delete the 5 redundant skill directories<br>   (`gemini-managed-agents-api`, `liveapi-service`, `documentation-and-adrs`, <br>   `gemini-api-dev`, `gemini-live-api-dev`) from `~/.gemini/config/skills/`.<br>2. **Generate Optimised `skills.json`**: Write the full-inventory <br>   `~/.gemini/config/skills.json` with 49 active skills (`//` prefix) and <br>   116 excluded skills.<br>3. **Verify Context Footprint**: Re-audit the active skill prompt token <br>   overhead to confirm successful reduction to ~3,350 tokens.<br><br>---<br>&gt; [!IMPORTANT]<br>&gt; **Mandatory Guardrail Confirmation**: In accordance with <br>  `organise-agent-skills` guardrails, please confirm if you would like me <br>  to proceed with deleting the 5 redundant skill directories and creating <br>  the optimised `~/.gemini/config/skills.json`.</pre><p>By trimming back to a core of 49 active skills and delegating the rest to meta-skills and on-demand rules:</p><ul><li>We slashed system prompt overhead by approximately 13,650 tokens; that’s an <strong>80.3% reduction</strong>.</li><li>We saved over <strong>400,000 tokens per pair-programming session</strong>.</li><li>Agent response latencies dropped noticeably.</li><li>Tool selection precision increased because the agent was no longer fighting distractor noise from over 100 overlapping skill descriptions. Instead of having a mild existential crisis every turn trying to choose between five different Gemini SDK guides, the agent simply picks the right tool and gets on with the job.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*S8oizfkz6ty5N_QaJU5_YA.png" /><figcaption>Optimisation results</figcaption></figure><h3>Summary &amp; Conclusion</h3><p>Agent skills are awesome. They’re so easy to find, install and use. But when we allow skills to sprawl unchecked, we start confusing our agents. This impacts accuracy, reliability, performance, and cost.</p><p>So increasingly, we need a way to manage and optimise our skills. For this, I’ve done some work so you don’t have to!</p><p>Feel free to download my optimisation skill off-the-shelf:</p><pre># Install all the Dazbo Agent Skills<br>npx skills add https://github.com/derailed-dash/dazbo-agent-skills<br><br># Install just the `Organise Agent Skills` skill<br>npx skills add https://github.com/derailed-dash/dazbo-agent-skills --skill organise-agent-skills</pre><p>Or if you have the Vercel find-skills skill installed, you can just ask your agent:</p><p><em>“Install </em><em>dazbo-agent-skills for me.&quot;</em></p><p>So, that’s it, folks. Go forth and optimise. If you find this skill useful, please give the <a href="https://github.com/derailed-dash/dazbo-agent-skills">repo</a> a star.</p><p><em>Have you experienced Skills Sprawl in your agentic environment? How many skills are currently active in your setup? Let me know in the comments below!</em></p><h3>Before You Go</h3><ul><li><strong>Please share</strong> 📢 this with anyone that you think will be interested. It might help them, and it really helps me!</li><li>Please <strong>give me loads of claps</strong>! 👏 (Just hold down the clap button.)</li><li>Please <strong>leave a comment</strong> 💬. Interaction is good!</li><li><strong>Add a star</strong> ⭐ on my repos!</li><li><strong>Follow</strong> 👉 and <strong>subscribe </strong>🔔, so you don’t miss my content.</li></ul><h3>References and Useful Links</h3><h4>Standards, Hubs &amp; Agent Platforms</h4><ul><li><a href="https://agentskills.io/home">Agent Skills Specification &amp; Open Standard</a></li><li><a href="https://skills.sh/">Skills Hub &amp; Directory (</a><a href="https://skills.sh/">skills.sh)</a></li><li><a href="https://antigravity.google/?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Google Antigravity</a></li></ul><h4>Related Dazbo Articles &amp; Repositories</h4><ul><li><a href="https://dazbo.co.uk/">Dazbo’s Portfolio</a></li><li><a href="https://dev.to/gde/dialling-our-agents-to-11-agent-skills-you-need-to-be-using-2o2k"><em>Dialling Our Agents to 11: Agent Skills You Need to be Using!</em></a></li><li><a href="https://medium.com/google-cloud/confused-about-where-to-put-your-agent-skills-ea778f3c64f3?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco"><em>Confused About Where to Put Your Agent Skills?</em></a></li><li><a href="https://medium.com/google-cloud/automated-github-code-reviewsusing-google-gemini-7b4d027b3092?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco"><em>Automated GitHub Code Reviews Using Google Gemini</em></a></li><li><a href="https://github.com/derailed-dash/dazbo-agent-skills">Dazbo Agent Skills Repository (</a><a href="https://github.com/derailed-dash/dazbo-agent-skills">derailed-dash/dazbo-agent-skills)</a></li></ul><h4>Key Agent Skills Repositories</h4><ul><li><a href="https://github.com/google/skills">Google Skills (</a><a href="https://github.com/google/skills">google/skills)</a></li><li><a href="https://github.com/google/agents-cli">Google Agents CLI &amp; Workflow Skills (</a><a href="https://github.com/google/agents-cli">google/agents-cli)</a></li><li><a href="https://github.com/google-gemini/gemini-skills">Google Gemini API Skills (</a><a href="https://github.com/google-gemini/gemini-skills">google-gemini/gemini-skills)</a></li><li><a href="https://github.com/addyosmani/agent-skills">Addy Osmani Agent Skills (</a><a href="https://github.com/addyosmani/agent-skills">addyosmani/agent-skills)</a></li><li><a href="https://github.com/Shubhamsaboo/awesome-llm-apps/tree/main/agent_skills">Awesome LLM Apps Skills (</a><a href="https://github.com/Shubhamsaboo/awesome-llm-apps/tree/main/agent_skills">shubhamsaboo/awesome-llm-apps)</a></li><li><a href="https://github.com/coreyhaines31/marketingskills">Corey Haines Marketing Skills (</a><a href="https://github.com/coreyhaines31/marketingskills">coreyhaines31/marketingskills)</a></li><li><a href="https://github.com/remotion-dev/skills">Remotion Video Skills (</a><a href="https://github.com/remotion-dev/skills">remotion-dev/skills)</a></li><li><a href="https://github.com/vercel-labs/skills">Vercel Labs Skills (</a><a href="https://github.com/vercel-labs/skills">vercel-labs/skills)</a></li></ul><h4>Academic Research &amp; Benchmarks</h4><ul><li>Paper: <a href="https://arxiv.org/pdf/2605.24660"><em>“How Many Tools Should an LLM Agent See? A Chance-Corrected Answer”</em></a> (Repantis et al., May 2026)</li><li><a href="https://gorilla.cs.berkeley.edu/leaderboard.html">Berkeley Function Calling Leaderboard (BFCL)</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=167f907e9dc2" width="1" height="1" alt=""><hr><p><a href="https://medium.com/google-cloud/skills-sprawl-when-too-much-of-a-good-thing-bloats-your-ai-agent-167f907e9dc2">Skills Sprawl: When Too Much of a Good Thing Confuses  Your AI Agent</a> was originally published in <a href="https://medium.com/google-cloud">Google Cloud - Community</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Finally — Hard Caps to Limit Your Google Cloud Spend]]></title>
            <link>https://medium.com/google-cloud/finally-hard-caps-to-limit-your-google-cloud-spend-29b2e658ed73?source=rss-58a0bd8ba726------2</link>
            <guid isPermaLink="false">https://medium.com/p/29b2e658ed73</guid>
            <category><![CDATA[gcp-billing]]></category>
            <category><![CDATA[finops]]></category>
            <category><![CDATA[kill-switch]]></category>
            <category><![CDATA[hard-cap]]></category>
            <category><![CDATA[google-cloud-platform]]></category>
            <dc:creator><![CDATA[Dazbo (Darren Lester)]]></dc:creator>
            <pubDate>Wed, 29 Jul 2026 09:49:50 GMT</pubDate>
            <atom:updated>2026-07-29T13:42:41.782Z</atom:updated>
            <content:encoded><![CDATA[<h3>Finally — Hard Caps to Limit Your Google Cloud Spend</h3><h3>TL;DR</h3><p>For years, Google Cloud budget alerts were purely informational. They sent an email while your project quietly burned through your savings.</p><p>I built my own Pub/Sub Killswitch to sleep peacefully at night, but it requires a bit of plumbing to install. What we really wanted was a easy option in the Google Cloud Console.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/515/0*iVRnfe5fulaoeeCu" /><figcaption>Good news everyone!</figcaption></figure><p>Good news! Google has finally introduced native Cloud Spend Caps (currently in Public Preview). You can now hard-cap spend per service so an unexpected API leak or runaway script hits a 403 Permission Denied wall instead of creating a $10,000 bill.</p><p>Here is how it works, how fast it triggers, and what you need to know…</p><h3>What’s the Problem?</h3><p>Over the last couple of years I’ve read weekly horror stories about individuals getting hit for thousands, or even tens of thousands of dollars of unexpected costs from Google Cloud.</p><p>Do a quick Google search, and you’ll find examples like…</p><ul><li><a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/google-cloud-customer-wakes-up-to-usd18-000-bill-despite-usd7-budget-thanks-to-forgotten-public-api-key-attacker-put-in-60-000-requests-and-blasted-through-usd1-400-spending-cap">“Google Cloud customer wakes up to $18,000+ bill despite $7 budget, thanks to forgotten API key in published project — attacker put in 60,000+ requests and blasted through $1,400 spending cap”</a> — Tom’s Hardware, April 2026</li><li><a href="https://www.reddit.com/r/googlecloud/comments/1pimbs9/got_hit_with_a_65000_bill_overnight_from_gcs_due/">Got hit with a $65,000 bill overnight from GCS due to a spike in list object calls</a> — Reddit, December 2025</li><li><a href="https://www.reddit.com/r/googlecloud/comments/1tr4vp8/google_cloud_billed_me_19000_usd_r105000_brl/">Google Cloud billed me ~$19,000 USD (~R$105,000 BRL) after an API key breach — and the charges keep growing even after I deleted everything </a>— Reddit, May 2026</li></ul><p>The big problem that most folks don’t understand is this:</p><p><strong>GOOGLE BUDGET ALERTS WERE NOT HARD LIMITS. THEY DID NOT PREVENT YOU FROM SPENDING MORE THAN YOUR BUDGET.</strong></p><p>Most of the examples I see follow the same pattern…</p><ol><li>A user follows Google best practice and sets a budget alert on their project. Say, $50 dollars per month.</li><li>They go to bed.</li><li>They wake up owing thousands.</li></ol><p>Reading these stories was causing me so much distress. Both for the many people getting caught out, but also because the same thing happening to me felt like an inevitable ticking time bomb.</p><p>In theory, I know more about Google Cloud than the average punter. I’m supposed to be a <a href="https://developers.google.com/community/experts?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Google Developer Expert (GDE)</a> and <a href="https://partners.cloud.google.com/google-cloud-ambassadors?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Google Ambassador</a>, after all!</p><p><em>But still I was terrified!</em> And I know at least a couple of fellow GDEs that have been hit with this. Which just goes to show: <strong>even if you know a lot about Google Cloud and think you’re doing everything right, it’s still possible to get caught out.</strong></p><p>And since Google didn’t offer any hard limit (aka <em>hard cap</em>) mechanism, there was literally no off-the-shelf solution.</p><h3>So I Built My Own Killswitch</h3><p>I built a mechanism that responds to budget alerts by detaching billing from whatever project(s) is/are associated with that alert. Sure, you can still get billing alert delays, but this mechanism typically stops any offending project inside of an hour of costs exceeding your budget. And this could be the difference that turns a £10,000 unexpected bill into a $100 unexpected bill. My solution is free and easy to install. And I blogged about it <a href="https://medium.com/google-cloud/how-to-avoid-a-massive-cloud-bill-41a76251caba">here</a>.</p><p>Using it requires a bit of setup…</p><ol><li>You need to create your budget alerts, of course.</li><li>You need those alerts to be sent to a Pub/Sub topic that you create.</li><li>You need to deploy a Cloud Run Function that does the actual work of responding to the alert and disconnecting a project.</li><li>You need to wire-up that Cloud Run Function such that it is triggered by events on the Pub/Sub topic.</li><li>You need pay careful attention to the roles you assign to the service account that runs this Cloud Run project — particularly if you happen to have many different Google Cloud projects.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ZGYyfj4UYtY5v_eZ.png" /><figcaption>My GCP Billing Killswitch</figcaption></figure><p>Even though the solution is well-documented and all of the above is scripted, it’s still a fair bit of work. It can be daunting for many.</p><p>I spoke to Google about this. A lot.</p><h3>Google Introduced Cloud Spend Caps!</h3><p>They delivered!</p><p>I tested this out in Private Preview over the last couple of weeks, but it went public today! So now y’all can use it too.</p><h3>Setting Up a Spend Cap</h3><p>First, <strong>create a Budget</strong> in the Google Cloud Console. Same place as before…</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/862/1*Zn57WfH2--0DDdm6hFXJHg.png" /><figcaption>Create budget alert</figcaption></figure><p>Traditionally, you would then set the amounts and then you’d see these available actions…</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/615/1*9ZcYbnHFE2PiC3yhFbjO9Q.png" /><figcaption>Traditional budget alert — no spend cap</figcaption></figure><p>But NOW you have the option to select <strong>“Spend cap enforcement”</strong> when you create the budget.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/572/1*ZEHEtyd5AHkxWb3uI3l0VQ.png" /><figcaption>Spend cap enforcement</figcaption></figure><p>If you select this, <strong>you MUST then specify a service</strong>. <em>(Currently you can only select one.)</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/569/1*p4A5Y8_Jv9qn0UCcYkJEiQ.png" /><figcaption>Specify a service</figcaption></figure><p>Note that if you use the “Spend cap” type, you can no longer select the option to publish to Pub/Sub:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/567/1*CmicBEemhV4GmxOEEeMQEw.png" /><figcaption>No Pub/Sub with Spend Caps</figcaption></figure><p>Once created, the new budget alert looks like this. Note how this new alert says<em> “Spend cap status: Configured”</em>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bNr2WLJGm2xpU26jsmFsIQ.png" /><figcaption>Spend cap configured</figcaption></figure><h3>Trying It Out</h3><p>I started running up some cost on my Gemini API inside this project.</p><p>Ooh, look! Inside my CI/CD pipeline I can see it’s hit a <strong>“403 — permission denied”</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WawRSGunGQjXmNPxl3cmog.png" /><figcaption>No permission!</figcaption></figure><p>And I got this email, almost immediately.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/933/1*SR2yM2X3u9E8C1_YaNnNBw.png" /><figcaption>Spend cap reached</figcaption></figure><p>It’s not too subtle! You’re gonna struggle to miss this in your inbox!!</p><p>We can take a look in Cloud Console to see more information:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oB4ajZpU0ep-JET4Py9rRQ.png" /><figcaption>Spend limit reached</figcaption></figure><p>Okay, successful test! It worked exactly as we wanted it to.</p><h3>How to Lift the Cap?</h3><p>What if we want to continue working with our service? In that case, we just need to lift the spend cap.</p><p>Open your budget and then click on <strong>“Lift spend cap” </strong>in the Console:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*C9Vw7cZURGXMMsqoLvmbmg.png" /><figcaption>Lift spend cap</figcaption></figure><p>You’ll see this message:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rdN6nSZNo0dowsp-9b2O7w.png" /><figcaption>Proceed?</figcaption></figure><p>Now you need to specify a new, <em>higher </em>amount:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*d5S8Hbaihpy-6JfGx0vTpA.png" /><figcaption>Lift in progress</figcaption></figure><p>You get a warning that it might take an hour to lift the cap. But for me, it took about 2 minutes:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gsOOzHJh0ZXW8gKg_cxx4g.png" /><figcaption>May take one hour</figcaption></figure><p>Then I received these two emails, nearly immediately:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/939/1*CG2SzF4JNQJ088Sx157CkA.png" /><figcaption>Email: spend cap lifted</figcaption></figure><p>Spend cap updated:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/934/1*C2hPUA7ezOhMbUELc8BAcg.png" /><figcaption>Email: spend cap updated</figcaption></figure><p>So when you’re working with spend caps, you always know exactly what’s going on.</p><h3>How Does This Differ From Spend Caps in AI Studio?</h3><p>It works the same way. But AI Studio only lets you set spend caps associated with Gemini API keys.</p><h3>The Most Frequently Asked Question…</h3><p>One of the biggest historical problems with budget alerts, and anything that leverages them (like my Billing Killswitch) is that the alerts are only triggered after billing cycle reconciliation. This results in a delay between your actual spend, and billing alerts firing. This delay could be hours. A lot of damage can be done to your credit card in a few hours!</p><p>With the new spend capping feature, Google promises that “near real-time enforcement”, with the <strong>caps triggering within minutes of actual spend thresholds being met</strong>. In my testing, it triggered within a couple of minutes.</p><p>Two minutes rather than two hours? I’ll take it!</p><p>Maybe, in the near future, we might expect some closer integration with the new Anomaly Detection feature. Wouldn’t it be nice if we had a toggle to enable capping to trigger based on the <em>prediction </em>of a spend cap being hit in the next few minutes or hours? (For the SREs out there… This would be a similar approach to setting alerts based on the rate of error budget consumption.)</p><p>Anyway, that’s my prediction. Let’s see!</p><h3>Spend Cap Limitations?</h3><p>Just a couple that I noticed.</p><ul><li>At the moment, you have to set these caps one service at a time. It would be better if we could pick multiple services. So for now, I’ll still be using my Killswitch mechanism when I want to limit billing across the whole project or collections of projects.</li><li>At the moment, you can’t have a cap and ALSO send a message to Pub/Sub. That might be useful for some.</li></ul><h3>Wrapping Up</h3><p>Native Cloud Spend Caps are an absolute game-changer. They provide the one thing cloud engineers have been begging for: actual, deterministic spend boundaries. If you’re building with high-throughput APIs like Gemini, get this configured immediately.</p><p>Good architecture isn’t just about high availability — it’s about keeping your wallet intact!</p><h3>Before You Go</h3><ul><li><strong>Please share</strong> this with anyone that you think will be interested. It might help them, and it really helps me!</li><li>Please <strong>give me loads of claps</strong>! (Just hold down the clap button.)</li><li>Please <strong>leave a comment</strong> 💬. Interaction is good!</li><li><strong>Add a star</strong> on the repo!</li><li><strong>Follow</strong> and <strong>subscribe, </strong>so you don’t miss my content.</li></ul><h3>Useful Links and References</h3><ul><li><a href="https://cloud.google.com/blog/topics/cost-management/new-early-anomalies-and-spend-caps-on-google-cloud-budgets?utm_campaign=DEVECO_GDEMembers&amp;utm_source=deveco">Google Cloud blog — Detect early and enforce firmly with Google Cloud’s enhanced cost controls for AI spend</a></li><li><a href="https://medium.google-cloud/how-to-avoid-a-massive-cloud-bill-41a76251caba">My GCP Billing Killswitch Blog</a></li><li><a href="https://github.com/derailed-dash/gcp-billing-killswitch">My GCP Billing Killswitch on GitHub</a></li><li><a href="https://dazbo.co.uk/">Dazbo’s Portfolio</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=29b2e658ed73" width="1" height="1" alt=""><hr><p><a href="https://medium.com/google-cloud/finally-hard-caps-to-limit-your-google-cloud-spend-29b2e658ed73">Finally — Hard Caps to Limit Your Google Cloud Spend</a> was originally published in <a href="https://medium.com/google-cloud">Google Cloud - Community</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>