AutoGen

The Python codebase is split into three layers. autogen-core is an event-driven, actor-style runtime where agents exchange messages and can run locally or across processes; it also has a .NET implementation. autogen-agentchat sits on top and provides ready-made agents (AssistantAgent, UserProxyAgent, CodeExecutorAgent) and teams. autogen-ext holds model clients, code executors and tool integrations such as MCP.

The main multi-agent model in AgentChat is a group chat: participants share one conversation and a policy decides who speaks next (fixed rotation, a model-based selector, handoff messages in Swarm, or the Magentic-One orchestrator). Termination conditions such as text mentions, token limits or handoffs end a run. GraphFlow adds an explicit directed-graph workflow on the same team interface.

State is held in memory during a run. Agents and teams expose save_state() and load_state() so an application can store and restore them, and the Memory protocol lets an agent pull stored items into its context. There is no built-in durable execution layer.

The AgentChat tutorial describes two approaches. During a run, a UserProxyAgent participates in the team and asks for input through a configurable input_func (console input() by default, or an async function wired to a web UI); while it waits, the whole team is blocked, so the docs advise using it only for short exchanges. Between runs, the team stops on a termination condition such as max_turns, TextMentionTermination or HandoffTermination, and the user's reply is passed as the next task; the team keeps its context. ExternalTermination and cancellation tokens let application code stop a run from outside. The docs link sample integrations for FastAPI, ChainLit and Streamlit.

pip install -U "autogen-agentchat" "autogen-ext[openai]"

import asyncio
from autogen_agentchat.agents import AssistantAgent
from autogen_agentchat.conditions import TextMentionTermination
from autogen_agentchat.teams import RoundRobinGroupChat
from autogen_ext.models.openai import OpenAIChatCompletionClient

async def main() -> None:
    client = OpenAIChatCompletionClient(model="gpt-4.1")  # reads OPENAI_API_KEY
    writer = AssistantAgent("writer", model_client=client,
                            system_message="Draft a two-line product blurb.")
    critic = AssistantAgent("critic", model_client=client,
                            system_message="Critique the draft. Say APPROVE when it is good.")
    team = RoundRobinGroupChat([writer, critic], max_turns=6,
                               termination_condition=TextMentionTermination("APPROVE"))
    result = await team.run(task="Blurb for a solar-powered kettle")
    print(result.messages[-1].content)
    await client.close()

asyncio.run(main())

• Python 3.10 or later is required.
• The README samples call the OpenAI API and expect OPENAI_API_KEY to be exported.
• The 0.4+ packages (autogen-agentchat, autogen-core, autogen-ext) are a rewrite; code written for AutoGen 0.2 must follow the migration guide.
• The project is in maintenance mode, so fixes are limited to bugs, security patches and docs; check the date of the latest release before depending on it.
• Only connect to trusted MCP servers; the README warns they can run commands locally.

Only in maintenance mode. The README says it gets no new features, is community managed, and accepts bug fixes, security patches and docs changes. New users are pointed to Microsoft Agent Framework.

Yes, as a client. McpWorkbench in autogen-ext connects an agent to an MCP server's tools. No A2A or AG-UI support was found in the official repo or docs.

GitHub reports CC-BY-4.0 because the main LICENSE file covers documentation. The code is under the MIT License in LICENSE-CODE.

AutoGen is Microsoft's repository, rewritten in 0.4 into Core, AgentChat and Extensions, now in maintenance mode. AG2 is a separate, community-run project that diverged from the AutoGen codebase in November 2024 and has since moved to its own async API.

Microsoft publishes an AutoGen to Microsoft Agent Framework migration guide and recommends that framework for new projects.