What AutoGen is and its status
AutoGen came out of Microsoft Research and has three parts. AgentChat builds conversational agents and teams, Core gives an event-based base for large systems, and Extensions connect models and outside services.
The key fact as of October 2026: the repository README announces maintenance mode. The project gets no new features, the community manages it, and contributions are accepted only for bug fixes, security and documentation.
Microsoft advises new users to use Microsoft Agent Framework, the successor to AutoGen. Its version 1.0 came out as ready for production, with long-term support and the A2A and MCP protocols for links with other agents and services. A migration guide exists for current projects.
AutoGen code is distributed under the MIT license and the documentation under CC-BY-4.0. Similar open tools are also MIT: CrewAI for agent teams with roles, and LangChain with LangGraph for chains and agents in Python and JavaScript.
Setup
AutoGen installs as a Python library. So from here you need a person who runs commands in a terminal and reads code: without that, you cannot build a team of agents. You need Python 3.10 or newer.
1. Create a virtual environment, a separate folder with the project’s libraries: python3 -m venv .venv, then source .venv/bin/activate (this is for Linux and macOS).
2. Install AgentChat with the command pip install -U “autogen-agentchat”.
3. Install a model client, the module through which the agents call the model. For an open model through Ollama it is pip install -U “autogen-ext[ollama]”, and for OpenAI cloud models pip install “autogen-ext[openai]”.
If you find code for AutoGen 0.2 in articles, you will have to update it: the README has a separate migration guide for code and settings from the old version.
The README example assumes an OpenAI cloud model with an API key. For a company in Russia, and for any company that keeps data in house, the practical path is different: an open model through Ollama on your own server, covered below.
First team: a writer and a critic
A team in AutoGen is a group of agents with a shared goal. The documentation walks through an example of two agents: one writes a text, the second criticizes and approves it. This technique is called reflection: the model checks its own work.
1. Create a model client and the first agent: an AssistantAgent named primary with the system_message parameter set to “You are a helpful AI assistant.”
2. Create a second agent, critic, with a system_message: give constructive feedback and reply APPROVE when the comments are addressed.
3. Set the stop condition TextMentionTermination(“APPROVE”): the team stops as soon as the critic writes this word.
4. Build a RoundRobinGroupChat team of the two agents with this condition in the termination_condition parameter. In such a team the agents speak in turn.
5. Run the team with the team.run method and a task; in the documentation it is a short poem about autumn. You can follow the work through Console, which prints the agents’ messages in the terminal.
For your own task, replace the poem with a task from your process, for example a draft reply to a client, and give the critic your rules: prices, terms and tone. The stop condition keeps the two agents from arguing forever.
Besides taking turns, AutoGen has three more ready patterns. SelectorGroupChat picks the next agent with the help of a model, MagenticOneGroupChat solves open tasks with web search and files, and Swarm passes the task between agents through an explicit handoff.
A local model through Ollama
Ollama runs open models on your computer or server and is distributed under the MIT license. AutoGen connects it through OllamaChatCompletionClient from the autogen-ext[ollama] package.
1. Start an Ollama server with the model you need; the beginner’s path is covered on the page about a local agent.
2. In the code, create a client: OllamaChatCompletionClient(model=”llama3.2”). The documentation example assumes Ollama runs locally on port 11434.
3. Pass this client to the agents from the example above in the model_client parameter instead of the cloud one.
The documentation marks Ollama support as experimental and warns that small local models are usually weaker than large cloud ones, and on some tasks the result may surprise you. For a team of agents this matters: a mistake by the first agent travels down the chain.
AutoGen Studio with no code
AutoGen Studio is a web interface where you build agents and teams without code. You install it with pip install -U “autogenstudio” and start it with autogenstudio ui --port 8080 --appdir ./my-app, and then it opens in the browser on port 8080.
The documentation calls Studio a research prototype that is not meant for production. It has no authorization, no checks against prompt injection and no other security features, and for running agent code the documentation advises a Docker container.
So Studio is good for showing a manager how agents pass a task to each other. Do not open it to employees, and do not give it real client data.
Where it breaks
The AutoGen documentation advises starting with one agent and moving to a team only when one is not enough. Anthropic gave the same advice in its review of agent systems on 19 December 2024: start with a simple solution and add complexity only when needed.
On 25 June 2025 Gartner predicted that more than 40% of agentic AI projects will be canceled by the end of 2027 because of costs, unclear value and weak risk control. A team of agents makes all three reasons stronger: more model calls, more places for an error, and it is harder to see which agent made the mistake.
The price also grows with the team. In the writer-and-critic example, each round of conversation is at least two model calls, so response time and model costs rise with the number of agents and rounds.
AutoGen has its own risk too: the project is in maintenance mode, and new features will appear in Microsoft Agent Framework instead. We build systems meant to last for years on a supported framework, with a log of every agent step.
When a team of agents is truly needed, for example when a request, a calculation and a contract go through different agents, it is work for a developer and the process owner. How such systems work is covered on the page about multi-agent systems, and we build a ready team of 3–5 agents as an AI department.