AutoGen

What it is

AutoGen is Microsoft’s framework for AI apps in which several agents work on a task, on their own or alongside a person. It comes in layers. The Core API passes messages between event-driven agents and can spread them across processes, in Python and .NET. AgentChat sits on top with preset agents and teams, and Extensions adds the model clients and code execution. AutoGen Studio is a no-code web app for building and running teams. Read the status first. The README now heads its caution “Project archived and no longer maintained”, and says Microsoft gives it no bug fixes, security fixes or support.

1. You pick a model client, such as OpenAI or Ollama
2. You give an AssistantAgent a name, a model and tools
3. You put agents in a team, or wrap them as tools
4. A termination condition decides when the run stops
5. Console streams each message, tool call and result

Who it suits

Python developers who already run AutoGen code and need it working while they plan a move, and people who want to study multi-agent patterns such as group chat, Swarm and Magentic-One. For a new project we’d follow the README and start on Microsoft Agent Framework, since the successor gets fixes and AutoGen no longer does.

Where to use it

AutoGen is a library and a local web app, not a plugin for a coding agent. The README and docs give no steps for Claude Code, Codex, Cursor or any other agent, so the kit follows the routes Microsoft documents.

• Python 3.10 or later: AgentChat and Extensions from PyPI
• AutoGen Studio: a web app on your own machine, on port 8080 unless you pick another
• .NET: pre-release NuGet packages for the Core API
• Magentic-One: a ready-made team for web and file tasks, which the docs say to run in Docker containers

Agents can also call MCP servers. The README’s second sample runs the Playwright MCP server through McpWorkbench.

How to set it up

Get it saves a zip of Mindket’s setup notes in your downloads folder, with every command taken from Microsoft’s README and docs. Python comes first, then Studio.

1. Make a virtual environment first. The installation page gives venv and conda steps, and the kit has the Windows lines.
2. Install AgentChat and the OpenAI client. AutoGen needs Python 3.10 or later.
3. Export your OpenAI key. The README’s samples read it from OPENAI_API_KEY, and setup/models-and-keys.md covers Azure, Anthropic, Ollama and Gemini.
4. Run the README’s Hello World sample. It creates one AssistantAgent and asks it to say Hello World.

python3 -m venv .venv
source .venv/bin/activate
pip install -U "autogen-agentchat" "autogen-ext[openai]"

For the no-code route, install AutoGen Studio and start it. The README’s line keeps Studio’s database and files in a folder you name.

pip install -U "autogenstudio"
autogenstudio ui --port 8080 --appdir ./my-app

How to use it

You write a short async Python script. Create a model client, give it to an AssistantAgent, and call run or run_stream with a task. The README’s multi-agent sample wraps two experts in AgentTool, so one general assistant calls them when a question needs them. For agents that talk in turns, the Teams page lists four presets.

The same page advises starting with one agent and its tools, and forming a team only when that agent falls short. GraphFlow, marked experimental, runs agents along a directed graph when the order must be strict.

Example

You want a first agent that uses a tool, so you run the Quickstart’s sample. It defines get_weather, a fake tool that always returns 73 degrees and sunny, and gives it to an AssistantAgent with reflect_on_tool_use on. The task it sends:

What is the weather in New York?

Console prints your message, then the agent’s FunctionCall to get_weather with the city New York, the tool’s result and the final reply. On the docs page that reply reads “The current weather in New York is 73 degrees and sunny.” Your model may word it differently.

For newcomers and experts

If you are new. An agent here is a language model with a name, instructions and tools. You give it a task, and it answers or calls a tool first. A team is several agents passing messages until a stop condition fires, such as a critic replying APPROVE in the Teams tutorial. Start with the Quickstart’s single agent, then read the Teams page. If you would rather click than code, AutoGen Studio builds the same teams by drag-and-drop, and its docs link a video tour.

For experts. The Core API is an actor-model runtime with event-driven agents that runs locally or across processes. Its gRPC packages let Python and .NET agents talk to each other. Teams, agents, model clients and termination conditions dump to JSON with dump_component(), and a team loads back with load_component(). That is how AutoGen Studio reads a team written in Python, and how a team built in Studio runs from code.

Logging goes through Python’s standard logging module, under autogen_core.EVENT_LOGGER_NAME with the event type LLMCall. ChatCompletionCache wraps a model client to cache its replies. Planning the move? Microsoft’s migration guide says AssistantAgent is single-turn unless you raise max_tool_iterations, while Agent Framework’s Agent keeps calling tools until it can answer. The guide still marks Anthropic and Ollama clients and the cache as planned there.

Good to know

Most of the risk sits in the status and in what agents can reach.

• Status: the README says Microsoft gives AutoGen no bug fixes, security fixes, new features or support, and advises against production use. GitHub’s API did not flag the repository as archived when Mindket read it
• AutoGen Studio: a research prototype, not a production app. Its security note says jailbreak tests and per-user limits on keys and data are not implemented
• Magentic-One: its agents browse, read files and run code. The docs say to use containers, watch the logs, keep a human in the loop and watch for prompt injection from web pages
• MCP servers: the README says to connect only trusted ones, since they may run commands on your machine or expose sensitive information
• Models: you bring your own key for a hosted model, or run a local one with Ollama. The Models page marks its Anthropic, Ollama, Gemini and Llama API sections as experimental
• The 0.2 line: Microsoft has no admin access to the pyautogen PyPI package any more, and its releases since 0.2.34 are not Microsoft’s. pip install autogen-agentchat~=0.2 keeps you on that line
• License: the docs and other content are under CC BY 4.0 and the code under MIT. Neither grants rights to Microsoft’s names, logos or trademarks