CrewAI

Mark your calendars and join!

🔥 CrewAI 0.121 is live This release includes major improvements across agent reasoning, guardrails, context handling, and system observability—focused on helping teams build more reliable, production-grade agent systems. What’s new in this release: • Reasoning field for Agents — Adds native support for capturing and surfacing LLM reasoning. • Initial Hallucination Guardrail — A foundation for detecting and handling hallucinated outputs (no-op implementation for now). • Markdown output support — Tasks now support structured markdown output natively. • Auto-injected timestamps — Agents can now dynamically receive contextual dates at runtime. • New tools and integrations — StagehandTool, MCP integration, and more • Documentation and bug fixes — Expanded docs, new examples, and broader support for event handling and logging. We’re continuing to push toward a future where multi-agent systems are robust, observable, and enterprise-ready out of the box. If you’re exploring agents in production, we’d love to hear what you’re building.

did you know you can customize the prompts in CrewAI? not just the task description, the actual system and user prompts sent to the model. it's not a "black box" as you might have read/heard. we do inject some helpful defaults to keep things structured and easy for those that don't need to fine tune further than we have. but if you want full control, you’ve got it. just write your own: ```Python from crewai import Agent agent = Agent( role="market analyst", goal="interpret economic data clearly", backstory="you explain trends with no jargon.", system_template="you are {role}. {backstory}\nyour goal is: {goal}", prompt_template="task: {input}\n\nrespond clearly.", use_system_prompt=True ) ``` or use a "custom_prompts.json" file to override just one slice: ```JSON { "slices": { "format": "thought: your reasoning\naction: tool name\ninput: tool input\nresult: your answer" } } ``` then wire it up: ```Python crew = Crew( agents=[agent], tasks=[task], prompt_file="custom_prompts.json" ) ``` want to see what’s actually being sent to the model? ```Python from crewai import Agent, Crew, Task from crewai.utilities.prompts import Prompts agent = Agent(...) task = Task(...) prompt_generator = Prompts(agent=agent, ...) generated_prompt = prompt_generator.task_execution() print("=== SYSTEM PROMPT ===") print(generated_prompt["system"]) print("=== USER PROMPT ===") print(generated_prompt["user"]) ``` more info in our docs → https[://]docs[.]crewai[.]com/guides/advanced/customizing-prompts

🔍 AI-Powered Lead Generator with CrewAI + Streamlit 🎈 Big shoutout to Albert Folch for shipping this clean and powerful multi-agent app for B2B prospecting—built entirely with CrewAI and Streamlit. Here’s what’s inside: 🧠 Lead Generator Agent – Finds relevant companies using Serper 🕵️♂️ Contact Agent – Scrapes key decision-makers across web + socials 📊 Lead Qualifier – Scores leads based on value + market signals 📋 Sales Manager – Generates a final, ready-to-use report Bonus: GPT-4o-mini token usage tracking for transparency—see exactly what each run costs. All wrapped in a slick Streamlit UI. Links in comments. Nice work, Albert! #CrewAI #AIagents #Streamlit #LeadGen

Our CEO João (Joe) Moura sat down with Aaron Fulkerson and Mark Hinkle on the AI Confidential podcast to go deep on the future of agents. they covered: ⚙️ Why building agents isn't the hard part — the real work is in measuring, evaluating, and integrating them 👩💼 Why the best people to train agents aren't devs, but ops managers 🔐 What “zero-trust” agentic infra actually looks like (and why it's the future) 🛠️ Why open-source orchestration frameworks like CrewAI matter more than ever Big thanks to the AI Confidential team for having us! Link to the full episode in the comments 👇

CrewAI 🤝 Google AI for Developers

we just shipped CrewAI v0.121.0! 🎉 so far, most of the agent have been jumping into accomplishing tasks without thinking. just like humans do. so yesterday we released agent reasoning. it's simple; agents now stop and plan before they act. the code change is literally one line: ```python reasoning=True ``` another bonus feature is automatically setting your task outputs to be in markdown. i joked at the webinar today i'll finally stop begging the agents to output in proper markdown 😂 my tasks used to look like this: ```yaml ... expected_output="IMPORTANT: make sure the output is in markdown format without the '```markdown' or '```' tags. ``` now it's just (set this in Task()): ```python markdown=True ``` more details in our docs! (https://docs[.]crewai[.]com)

Is MCP all you need ? MCP allows you to connect an LLM to your external tools and data. I'm particularly bullish on the idea that more and more services will enable native MCP servers. Think Salesforce, Gmail, Slack and Hubspot native MCP servers. Now your agents or LLM's can seamlessly connect to them. So are MCP's all you need ? Slides through my take👇

CrewAI 🤝 MCP

Day 3 of our CrewAI Launch week! 🎉