AutoGen
AutoGen is an open-source Microsoft framework for building apps where several AI agents talk to each other, and sometimes to people, to finish a task.
AutoGen started as a Microsoft Research project. It is an open-source toolkit for building programs out of several AI agents. An agent here is a program that takes part in the conversation; many run on a language model and can call tools. In AutoGen, agents solve a task by talking to each other in a shared conversation. A person can join that conversation too.
The main building block for beginners is called AgentChat. It gives you ready-made agents. One kind is an assistant that runs on a language model and can call tools. A tool is an extra ability the agent can use. Another kind of agent runs code. You group agents into a team. In a round-robin team, agents take turns in a fixed order, and each one sees everything the others said. The team keeps going until a stop rule, called a termination condition, says the job is done. AgentChat sits on a lower layer called Core, which handles the messages passed between agents.
One important fact before you pick it: AutoGen has moved into maintenance mode. That means it will not get new features. The community looks after it now. Microsoft points new users to its successor, Microsoft Agent Framework, which the AutoGen team helped build.
One agent working alone is not always enough for a hard task.
Follow one writing task through a two-agent team.
- 1 · agentsYou create agents, each with a model and its own instructions, such as a writer and a critic.
- 2 · teamYou put the agents in a team, for example one where they take turns in a fixed order.
- 3 · talkEach agent, on its turn, sends its reply to every other agent, so they all share the same conversation.
- 4 · checkAfter each reply, the team runs its termination condition, such as looking for the word APPROVE.
- 5 · resultWhen the condition is met, the run stops and returns a task result with the messages.
An AutoGen team is a conversation with a stop rule.
| Who | What they ask | What it works with |
|---|---|---|
| Student project team | “Can a second agent check the first one's essay draft?” | A writer agent and a critic agent taking turns |
| Developer building a coding helper | “Can one agent write code and another run it?” | An assistant agent paired with a code executor agent |
| Team building an app around agents | “How can our app stop a run from outside it?” | An external termination condition |
| Existing AutoGen user | “Where should our next project start?” | The migration guide to Microsoft Agent Framework |
- It gives you ready-made agents and team patterns, so you do not start from zero.
- It lets agents work together in one shared conversation.
- It lets a person join in, through an agent that passes user input to the team.
- It offers built-in stop rules, such as a message limit, a time limit or a keyword.
- It is in maintenance mode, so it will not get new features.
- Its own README tells new users to start with Microsoft Agent Framework instead.
- Teams need more steering than a single agent, so the docs advise starting with one agent for simpler tasks.
- Its no-code tool, AutoGen Studio, is not meant to be a production-ready app.
Sources used
This explainer is written in original language. The links below support its factual claims.
- repomicrosoft/autogen, Microsoft (GitHub) · read 28 Sept 2026
- paperAutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation, Wu et al., arXiv · read 28 Sept 2026
- docsAgentChat, Microsoft · read 28 Sept 2026
- docsAgents, Microsoft · read 28 Sept 2026
- docsTeams, Microsoft · read 28 Sept 2026
- docsTermination, Microsoft · read 28 Sept 2026
- docsAutoGen to Microsoft Agent Framework Migration Guide, Microsoft Learn · read 28 Sept 2026