AutoGen (Microsoft)
Most frameworks ask you to lay out an agent's control flow in advance. AutoGen, the open-source agent framework from Microsoft Research, asks for something different: define a handful of agents, let them send each other messages, and let the conversation itself become the control flow.
How the pieces fit
Everything in AutoGen is a conversable agent — an addressable participant with its own model, instructions and tools:
- An assistant agent reasons, writes code and calls tools.
- A user proxy stands in for the human: it runs the code the assistant produced, usually inside a container, and can pause for a person's approval.
- A group chat puts several agents in one room, with a manager choosing who speaks next.
A task ends when a termination condition fires — a keyword in a reply, a round limit, or a token budget.
Where it fits
The conversational framing suits work that genuinely benefits from critique: a coder paired with a reviewer, a planner challenged by a sceptic. Built-in code execution makes it strong on data analysis and engineering tasks. The trade is predictability — a free-form chat can loop, drift, or stall with agents politely thanking each other, which is why later versions moved to an asynchronous, event-driven core with explicit termination handling. Reach for LangGraph when you want a graph you can read; reach for AutoGen when you want agents to argue their way to an answer.
AutoGen is Microsoft's open-source framework for building multi-agent LLM applications as conversations: developers define specialised agents — an assistant, a code executor, a human proxy — and AutoGen routes messages between them until the task is done. Its distinguishing choice is treating agent collaboration as chat, with configurable termination and group-chat managers, which makes delegation, code execution and human approval composable rather than hand-wired into one prompt loop.
Imagine handing a job to a small team instead of one person. Someone drafts, someone checks the work, someone actually runs it — and they sort it out between themselves by talking.
AutoGen is a free toolkit from Microsoft for building software that works this way. Instead of writing out every step yourself, you create a few AI assistants, give each one a role, and let them exchange messages until the task is done. One of them can run code; another can stop and ask you before anything risky happens.
It suits jobs where a second opinion genuinely helps — writing and reviewing, calculating and double-checking.