# Microsoft AutoGen — Microsoft’s open-source framework for orchestrating conversations between multiple AI agents.

> Source: The Agents Index — https://theagentsindex.com/microsoft-autogen (structured, researched, re-verified)
> Facts last verified: 2026-07-24

AutoGen is Microsoft’s open-source framework for building multi-agent AI applications, systems where several LLM-powered agents converse, delegate and collaborate to complete a task, with or without a human in the loop. Released under the MIT licence and written primarily in Python (with a parallel .NET implementation), it pioneered the “agents talking to agents” pattern and remains one of the most-starred agent frameworks on GitHub (59,867 stars as of 2026-07-21) even though it is now in maintenance mode. Microsoft’s successor, the Microsoft Agent Framework, is no longer just an announced plan, it shipped as a production-ready 1.0 GA release on 2026-04-03, unifying AutoGen’s orchestration with Semantic Kernel’s enterprise plumbing under one supported stack for .NET and Python. The original AutoGen authors separately maintain a genuinely active community fork called AG2, which still controls the old PyPI packages.

| Fact | Value |
| --- | --- |
| Website | https://microsoft.github.io/autogen/ |
| Best for | Python or .NET developers and researchers exploring multi-agent orchestration who want a mature, free, self-hosted framework. |
| Not for | Teams starting a new long-term production build who need active feature development, Microsoft itself points them to the Agent Framework. |

## Verdict

AutoGen is the framework that popularised the “agents conversing with agents” model, and it remains a clean, well-documented and genuinely free way to prototype multi-agent systems in Python or .NET. Its layered design is a real strength: you can start in the high-level AgentChat API and drop into the Core runtime when you need fine-grained control over message passing. The catch is strategic, not technical, and it has hardened since AutoGen’s own SEED write-up: Microsoft’s successor, the Microsoft Agent Framework, is no longer just announced, it reached a production-ready 1.0 GA on 2026-04-03 and is the live recommended path for new projects, while the original authors’ AG2 fork remains genuinely active in parallel (shipping commits within the last week) and controls the old PyPI packages. AutoGen itself hasn’t tagged a release since September 2025. That three-way split is a real, current risk for anyone building something they must maintain for years, and multi-agent loops can be token-heavy and hard to make deterministic. Pick AutoGen today to learn the patterns or extend an existing AutoGen codebase; if you are starting fresh for production, evaluate the Microsoft Agent Framework or AG2 first.

## Strengths and weaknesses

- ✓ Mature, well-documented multi-agent conversation patterns with a large community, still the most-starred name in the space (59,867 GitHub stars) even though new stars now mostly flow to its successors.
- ✓ Layered API, prototype quickly in AgentChat, drop to Core for control; .NET support alongside Python.
- ✗ Confirmed frozen, not just announced: the latest tagged release is python-v0.7.5 (2025-09-30, ~10 months stale) and the most recent commit (2026-04-06) only updated the README’s maintenance banner. Microsoft’s official successor, the Microsoft Agent Framework, already reached a production-ready 1.0 GA on 2026-04-03, it is the live recommended path today, not a future plan, while the original authors’ AG2 fork is genuinely active in parallel (commits as recently as 2026-07-20). The three-way AutoGen / AG2 / Agent Framework split is a real, current adoption risk, not a theoretical one.
- ✗ Multi-agent conversation loops can be token-hungry and hard to make deterministic, and the layered concepts are heavier than a single-agent library.
- ⚠ It is a code library, not a hosted service, you build, host and operate the agents yourself.
- ⚠ Feature-frozen: for long-lived projects plan a migration path to the Microsoft Agent Framework or AG2.

## Key features

- **AgentChat multi-agent patterns** — A high-level API with prebuilt patterns, two-agent chat, group chat and round-robin teams, for rapid multi-agent prototyping.
- **Event-driven Core runtime** — A low-level message-passing API with a local or distributed runtime for fine-grained control over how agents communicate.
- **Extensions & model clients** — Pluggable model clients (OpenAI, Azure OpenAI and others) and tool integrations that expand what agents can do.
- **AutoGen Studio** — A no-code GUI for building and testing multi-agent teams visually, without writing framework code.

## Use cases

- **Multi-agent code generation** — A writer agent produces code and an executor agent runs and critiques it in a loop until the task passes.
- **Research & critique teams** — Several specialist agents draft, review and refine an analysis, with a planner routing subtasks between them.

## Integrations

OpenAI & Azure OpenAI model clients · Python function/tool calling · .NET / C# runtime · AutoGen Studio (no-code GUI) · Model Context Protocol (MCP) — the autogen-ext package connects agents to external MCP servers as a client (STDIO/SSE/Streamable HTTP)

## Sources

- https://github.com/microsoft/autogen
- https://microsoft.github.io/autogen/
- https://alexbevi.com/blog/2026/06/18/two-lineages-one-framework-how-autogen-and-semantic-kernel-became-the-microsoft-agent-framework/
- https://devblogs.microsoft.com/foundry/microsoft-agent-framework-reaches-release-candidate/
- https://visualstudiomagazine.com/articles/2026/04/06/microsoft-ships-production-ready-agent-framework-1-0-for-net-and-python.aspx
- https://github.com/ag2ai/ag2
- https://microsoft.github.io/autogen/stable//reference/python/autogen_ext.tools.mcp.html