crewAIInc/crewAI

CrewAI is a Python framework for building multi-agent automations. Its main fit is developers who need more than a single prompt: several specialized agents that use tools and hand work to each other, plus deterministic code around them.

Two ways to structure work

The framework has two complementary primitives. Crews model a team of agents, each with a role, goal, backstory and tools, and let them collaborate autonomously — for example a researcher and a reporting analyst passing findings along a sequential process, or a hierarchical process where a manager agent delegates. Flows are event-driven workflows with typed shared state, branching and routing. A Flow can start, call a Crew as one step, inspect the result, and route to different follow-up steps (Flows and Crews overview, combined example). That combination is the distinguishing mechanism: autonomy where judgment helps, explicit Python control where business logic must be reliable.

A concrete workflow

A typical project is created with crewai create crew <name>, which scaffolds agent definitions in agents/*.jsonc, tasks and process settings in crew.jsonc, plus tools/, knowledge/ and skills/ folders. The generated configuration wires a researcher using web search to an analyst that writes a markdown report to output/report.md. Users then run crewai install and crewai run (getting started, running a crew). An older Python/YAML scaffold is available with --classic. Model choice is configurable per agent, and the docs cover hosted providers as well as local options such as Ollama and LM Studio (LLM connections).

Requirements and practical constraints

The documented path requires Python 3.10 up to but not including 3.14, and uses the uv package manager to install the CLI; a full walkthrough is in the installation guide. Running agents needs a model provider API key, and the web-search example additionally needs a Serper.dev key. On Windows, a known chroma-hnswlib build error is addressed by installing Visual Studio Build Tools with the C++ desktop workload. The README also points coding agents at optional CrewAI Skills for Claude Code and other agents — this is a helper for using the framework, not a runtime dependency (Build with AI).

Data and licensing

CrewAI is MIT-licensed (license). It sends anonymous telemetry about versions, OS, agent/task counts and model names; prompts, task text, tool calls, responses and secrets are not collected unless a user explicitly enables share_crew, which then includes goals, backstories, task context and outputs. Telemetry can be disabled via OTEL_SDK_DISABLED (telemetry). A separate commercial AMP Suite adds managed deployment, tracing and enterprise support; the open-source framework does not require it.

One HN commenter describes running agents in production and contributing time-awareness and ID-generation tools to counter date hallucinations. That is a single unverified account, but it illustrates the kind of coordination detail production agent systems still need to handle.