CrewAI

Multi-agent work in this Python framework splits into two building blocks. Crews group agents, each given a role, goal, backstory, LLM and toolset, and run their tasks in order or under an auto-assigned manager that delegates and reviews results; Flows add event-driven control with shared state, branching and routing, and can call a crew as one step. It stands alone with its own primitives rather than wrapping another agent library, and adds memory, knowledge files, guardrails, checkpoints, async runs, plus MCP and A2A connectivity. A CLI installed through uv scaffolds projects where agents and tasks live in JSONC files, with the older YAML layout still available, then runs them with crewai run. Models default to OpenAI, but other providers or a local Ollama or LM Studio server can be configured. It needs Python 3.10 to 3.13. The code is MIT licensed and free for commercial use; the company sells an optional AMP control plane for hosted deployment, tracing and governance. Anonymous telemetry is on by default and switches off with CREWAI_DISABLE_TELEMETRY. A sensible choice for developers building role-based agent pipelines.