How to Use AutoGen

Microsoft AutoGen is an open-source framework for building multi-agent LLM applications that solve problems through conversation.

AutoGen treats AI workflows as conversations between specialized agents. Instead of chaining prompts, you define an AssistantAgent that can write code, a UserProxyAgent that can execute it, and optionally a group-chat manager that routes turns between many agents. The loop continues until a termination condition is met — a regex match, a max-turn limit, or a human “exit”. The result feels less like a pipeline and more like a team meeting.

What it is

AutoGen is maintained by Microsoft Research and released under the MIT license. The current line is AutoGen 0.4+, which splits into three packages: autogen-core (event-driven runtime), autogen-agentchat (the high-level chat API most users touch), and autogen-ext (model clients and tools). Python is the primary language; a .NET port exists.

Install

pip install -U "autogen-agentchat" "autogen-ext[openai]"
# or for the classic 0.2 API
pip install pyautogen

First run

A minimal two-agent assistant that can call tools and stop when it says TERMINATE:

import asyncio
from autogen_agentchat.agents import AssistantAgent
from autogen_ext.models.openai import OpenAIChatCompletionClient

async def main():
    model = OpenAIChatCompletionClient(model="gpt-4o-mini")
    agent = AssistantAgent("writer", model_client=model)
    result = await agent.run(task="Draft a 3-line product tagline.")
    print(result.messages[-1].content)

asyncio.run(main())

Everyday workflows

• Pair an AssistantAgent with a UserProxyAgent that has code_execution_config set to run Python in Docker.
• Build a RoundRobinGroupChat of planner, coder, and critic agents for harder tasks.
• Wrap the whole thing with AutoGen Studio’s web UI to prototype without writing glue code.

Gotchas and tips

The 0.2 to 0.4 jump was a full rewrite — tutorials you find on blogs may import names that no longer exist. Pin to one major line per project. Also watch termination: without a TextMentionTermination or MaxMessageTermination, agents can loop politely forever and burn your token budget.

Code execution is powerful but dangerous. Always point UserProxyAgent’s executor at a sandboxed Docker image, not the host shell, and set a working directory outside your repo root so models cannot clobber your code.

Who it’s for

Researchers and backend engineers exploring agentic patterns beyond a single LLM call. Tip: start with two agents; only add a third when you can name a concrete reason the pair cannot solve the task.