Agent Framework
microsoft/autogen
microsoft/autogen· Python
Microsoft's multi-agent conversation framework for orchestrating LLM agents that talk to each other.
GitHub stats
- Stars
- 57,578
- Forks
- 8,681
- Watchers
- 517
- Open issues
- 792
meta
- License
- CC-BY-4.0
- Primary language
- Python
- Last commit
- 2026-04-15
- Stats fetched at
- 2026-04-29
AutoGen is Microsoft Research's Python framework for building agentic AI systems where multiple LLM agents converse to solve tasks. The v0.4 rewrite introduces a layered architecture: a low-level async messaging core, a high-level AgentChat API for typical multi-agent patterns (RoundRobin, Selector, Swarm), and AutoGen Studio for no-code prototyping. Install via `pip install autogen-agentchat autogen-ext`, then define agents with roles, tools, and termination conditions. Common uses include code generation with executor agents, research workflows, and human-in-the-loop assistants.
Editor's verdict
Pick AutoGen when your problem genuinely fits multi-agent conversation — e.g. coder + reviewer + executor loops, or you need Microsoft's research-backed patterns. The v0.4 async core is much cleaner than v0.2, but the API churn means a lot of tutorials online are outdated, and debugging multi-agent chats is still painful. For single-agent tool use, LangGraph or plain OpenAI SDK + a state machine will be simpler; for production-grade tracing and durability, look at LangGraph or Temporal-based stacks.