What is Microsoft AutoGen, and what problems does it solve compared to single-agent chatbots?

What you need to know

Where a single chatbot struggles

A single agent has to carry everything in one prompt: plan the work, do it, check it, call every tool. Three problems follow:

• Instructions dilute each other. A 3,000-word system prompt that says "write code", "review code" and "never touch production" gets followed unevenly.
• No outside check. A model reviewing its own answer in the same context tends to agree with itself.
• No grounding. The model can say its code works; nobody actually ran it.

What AutoGen adds

• Role split. Each agent gets a short, focused prompt and only the tools its job needs.
• Conversation as control flow. A team (group chat) decides who speaks next, so a loop such as write, run, read the error, fix can repeat until a stop condition is met.
• Code execution. A code executor agent runs model-written code in Docker or another sandbox and returns the real output or traceback.
• Humans as participants. A UserProxyAgent puts a person into the same conversation for input or approval.

The versions you must keep apart

A minimal 0.4+ program (pinned to autogen-agentchat 0.7.x):

1import asyncio2from autogen_agentchat.agents import AssistantAgent3from autogen_ext.models.openai import OpenAIChatCompletionClient45async def main() -> None:6    client = OpenAIChatCompletionClient(model="gpt-4.1")7    agent = AssistantAgent("assistant", model_client=client)8    result = await agent.run(task="Explain UPI in two sentences.")9    print(result.messages[-1].content)10    await client.close()1112asyncio.run(main())

Everything is async. run returns a TaskResult holding every message produced, and run_stream yields them as they happen.

When not to use it

Multi-agent means more LLM calls, more latency and a new failure mode: agents talking in circles. If one agent with two or three tools does the job, that is cheaper and easier to debug.

A real-life example

A Bengaluru fintech team wants help reviewing pull requests. Version one is a single chatbot: paste a diff, get comments. It praises code that does not even import correctly, because it never runs anything.

Version two is a code-review agent pair plus an executor:

• author writes or patches the function.
• executor (a CodeExecutorAgent in a Docker container) runs the unit tests.
• reviewer reads the diff and the test output, and replies APPROVED or at most three numbered issues.

On 50 old pull requests, the single chatbot missed 14 known bugs; the team of three missed 5, mostly because the tests really ran. Each review now costs about four model calls instead of one, so the team only uses it for pull requests that touch payment code.

Follow-up questions to expect

• "Is AutoGen still maintained?" — Microsoft's AutoGen is in maintenance mode: bug fixes and security patches, no new features. Microsoft recommends Microsoft Agent Framework for new projects and publishes a migration guide.
• "What is AG2?" — A community fork of AutoGen that continued the 0.2-style API under open governance. It is a separate project; the old pyautogen and autogen package names on PyPI belong to it, not to Microsoft.
• "What is AutoGen Studio?" — A low-code web UI on top of AgentChat for prototyping teams. It is for experiments, not production.
• "What is Magentic-One?" — A generalist multi-agent system built on AutoGen, where an orchestrator plans and directs web, file and code agents; it ships as MagenticOneGroupChat.