Arc Intelligence

Last night we hosted a hackathon with HackerSquad by Developer Events at betaworks! Huge shout-out to Brian Matzelle for building something that perfectly captures what we're excited about: making agents smarter over time at inference time i.e. getting better on the job. Here's what Brian built with our SDK: https://lnkd.in/erSdNwY4 He had an agent with 10 MCP tools (search, edit, analyze, etc.) and wanted it to get better at tool selection and learn which tools work best for different tasks, without constant retraining. ATLAS wrapped his MCP agent, judged tool efficacy via our reward system, and built learning playbooks from successful patterns. The agent automatically preferred high-reward tool sequences and received Teacher guidance when deviating from proven approaches. Shout-out to Adam Chan for bringing it all together 🚀

We're excited to be participating in Hack Night at Betaworks tonight! If you're in the neighbourhood, come say hi 👋

At Arc, similar to other folks at the frontier research of AI, we have been thinking deeply about continual learning. We, too, have followed the research and capabilities of reinforcement learning, continual learning, and the implications for world models. Over the past five months, our team has been particularly obsessed about the following question: What does it actually mean for AI agents to genuinely learn and improve over time? As a result, we’ve explored what might it look like for agents to adapt and learn at inference-time. Instead of asking "How do we update weights without forgetting?" We ask: "Can we skip weight updates entirely?" What did we discover? When we approach continual learning and adaptation as an orchestration problem, we are able to dynamically adjust supervision levels, retrieve distilled experience, and coordinate dual-agent handoffs based on task history. No gradients, no retraining, no specialized hardware required. We evaluated our system on a real cyber-threat investigation benchmark by Microsoft, and proved SOTA performance: 1️⃣ 54.1% task success with GPT-5-mini vs GPT-5 (High)'s 48.0% 2️⃣ 86% lower cost per task 3️⃣ Cross-incident transfer: Frozen artifacts from one scenario improve accuracy 28% → 41% on new scenarios with zero retraining We’re super proud of our approach to continual learning, the results of our benchmark, the open-source system we designed, and what’s next. We’d love for you to read the full paper, check out our blog, and try out the code in the comments below.

Getting started with the Atlas SDK? Check out our new demo video 👇 Atlas is a drop-in learning harness that enables your agent to learn from experience, adapt to new challenges, and become more efficient over time - all without modifying your existing agent code. We put the SDK to the test with a supply chain agent, and the results demonstrate a clear "J-curve" of learning. ▶️ How it Works: Atlas wraps your existing agent with an adaptive dual-agent reasoning loop guided by reward signals. This lets agents stay fast on familiar work while escalating supervision on new or risky tasks. The SDK records rich telemetry from these interactions, allowing the agent to learn and adapt over time. ▶️ The Results: The impact is immediate and measurable. After an initial learning dip, the agent's runtime efficiency skyrocketed. By scenario 3, it achieved a +52.8% improvement. After just 5 scenarios, it was running +25.8% more efficiently than its baseline. This translates to tangible outcomes: what was once a verbose, unstructured report from a frontier model becomes a perfectly structured, efficient JSON output. The agent gets faster, cheaper, and more effective with every interaction. Explore the Atlas SDK 👇 GitHub: https://lnkd.in/erSdNwY4 Documentation: https://docs.arc.computer/ #ContinualLearning #Agents

👉 Kicking off our first public release notes for the Atlas SDK, which we'll be sharing regularly. This week: v0.1.3, a continual-learning harness for production agents that turns every task into a structured learning episode. Full post: https://lnkd.in/echUe3Zw Technical Highlights for v0.1.3: 1️⃣ Adaptive Runtime: A core runtime routes each request through a capability probe that selects one of four execution modes: auto, paired, coach, or escalate. This architecture balances autonomous execution with graduated supervision, ensuring low latency on familiar tasks while capturing rich data on complex ones. 2️⃣ Persistent Learning & Memory: The SDK now includes a persistent memory system that captures agent instructions and human guidance from each run, tagged by reward signals. This compounding knowledge base can be automatically reused in future tasks. An optional Postgres backend is supported for production. 3️⃣ Structured Data for Training: Full learning trajectories, including plans, step attempts, guidance notes, and reward payloads, can be exported as structured JSONL. This provides clean, telemetry-rich data ready for evals or training pipelines. 4️⃣ Compatibility: The atlas CLI provides scaffolding for triage adapters (atlas triage init). Agents can be configured via YAML with support for Python, HTTP, and OpenAI-compatible endpoints. The full source code, architecture diagrams, and technical documentation are available on GitHub 👉 https://lnkd.in/erSdNwY4

ATLAS SDK is now LIVE: the learning harness for your agents. We enable your agents to learn and adapt in real-time on your most critical, long context workflows. Add our harness to boost performance, reduce token costs and build persistent memory. All you need to do is BYOA: Bring Your Own Agent and your agents evolve from static to dynamic, instead of just executing tasks, they build skills and get better with every run. Here's how to get started: 1. pip install arc-atlas 2. BYOA: plug in your agent (config file) 3. Define what you want to optimize for (rewards) 4. Execute your workflow Arc is fully open-source and deployable on your own infrastructure, giving you complete control over your environment. You own everything, the code, the data, and the skills your system develops and can export datasets at any time for fine-tuning or transferring knowledge elsewhere. A managed cloud version is on the way, so stay tuned. We want to see what you build. Our team is ready to forward-deploy to help you integrate and ship. Try SDK: https://lnkd.in/erSdNwY4 Docs: https://lnkd.in/eFZ6hD4V Talk to us: agent@arc.computer

What is continual learning? How is it different from reinforcement learning? Why does it matter for the future of AI systems? Reinforcement Learning optimizes actions in a fixed world, while Continual Learning adapts the model when the world changes. This shift from policy optimization to model adaptation is the key to building agents that can handle dynamic, real-world environments. See our deep dive into the key terms, what matters, and why CL is so critical for creating truly autonomous systems. If you are building or conducting research in this space, would love to connect! 👉 Full Blog: https://lnkd.in/eWpjzrqb

Join us at Cloudflare on Friday, October 3rd! RSVP: https://lnkd.in/eQBe8jnk