AutoGen: Conversational Multi-Agent Patterns and Their Successor

Remember the basic agent from earlier topics: one model in a loop — think, call a tool, read the result, think again (Topic 269). It works until the task gets big. Then a single loop starts to hurt:

• One context window is trying to hold everything — the plan, the code, the test output, the math. The desk overflows (Topic 278).
• The model grading its own work is a wolf checking on the sheep: it rarely disagrees with itself.
• A coding task needs a place to run code — a capability no text model has.

So a strange idea, from Microsoft Research in late 2023, flipped the design:

Instead of one super-agent with tools, you put several roles in a room — one that writes code, one that runs it, one that checks it, a human proxy who approves — and let the task get solved through messages between them. That idea, formalized, is AutoGen.