AutoGen

Sweet spot: a developer or researcher who genuinely needs multi-agent collaboration — not because it's trendy, but because the task decomposes cleanly into specialised roles. Code generation with a critic loop is the canonical AutoGen win. Failure modes. Teams reach for multi-agent when a single agent with good tool-use would do the job faster and cheaper. AutoGen's abstractions only pay off past a real complexity threshold — below it, the overhead is just noise. The 0.4 rewrite also means you should treat anything written for 0.2 as historical context, not a working template. What to pilot. Pick one task where you can articulate the agents' distinct roles (e.g., "planner decomposes, coder writes, tester validates"). Build it end-to-end in AutoGen, then build the same thing as a single-agent-with-tools approach. Compare output quality, cost, and debuggability. If multi-agent wins on at least two of those three, commit; if not, the simpler approach will scale better.

AutoGen is Microsoft Research's open-source framework for orchestrating multiple LLM agents that talk to each other to solve a task. Rather than writing a single long prompt, you define agents — a "Planner," a "Coder," a "Critic," a "UserProxy" — give each one its own system prompt and tool access, and let them converse. AutoGen handles the message routing, tool-use loops, code execution, and termination conditions. The current 0.4 rewrite (late 2024) split the framework into layers: a lightweight core event system, an agent API, and higher-level "teams" abstractions (RoundRobinGroupChat, SelectorGroupChat, etc). There is also AutoGen Studio — a no-code UI to prototype multi-agent workflows before dropping into Python. AutoGen is one of the most-cited multi-agent frameworks in academic work and is used in Microsoft's own products. It is MIT-licensed, model-agnostic (OpenAI, Anthropic, Azure, local), and has a large ecosystem of example agents and patterns.