Creativity · Agent Protocol
AutoGen GroupChat
AutoGen, Microsoft Research's multi-agent framework, provides a primitive called GroupChat. Several agents — each with its own role and prompt — sit around a virtual table. A GroupChatManager picks the next speaker based on the conversation so far; the chosen agent contributes; the cycle repeats until a termination condition. The pattern scales well to 3-10 agents and is especially strong for brainstorming, debate, and review workflows.
Protocol facts
- Sponsor
- Microsoft Research
- Status
- stable
- Spec
- https://microsoft.github.io/autogen/
- Interop with
- Azure OpenAI, OpenAI, A2A, MCP
Frequently asked questions
How does the GroupChatManager pick the next speaker?
Typically via an LLM call — it reads the transcript and returns the name of the next agent to speak. AutoGen also supports deterministic policies like round-robin, or user-defined selection functions.
When is GroupChat better than orchestrator-worker?
When the conversation is more organic than hierarchical — you don't want one boss planning, you want several experts chiming in when relevant. Code review with a security reviewer, a performance reviewer, and a style reviewer is a good example.
What's the risk with GroupChat?
Cost and drift. Many agents × many turns × long context burns tokens fast, and without careful prompting they can start echoing each other. Termination rules and a tight agent roster matter.
Sources
- AutoGen — Microsoft Research — accessed 2026-04-20