AutoGen

AutoGen popularized a genuinely useful idea for multi-agent systems: model a task as a group of conversable agents (an assistant, a code executor, a user proxy, a critic, etc.) that pass messages to each other, with the framework handling turn-taking and termination conditions. The rewritten v0.4 architecture (autogen-core + autogen-agentchat + autogen-ext) is an event-driven, actor-style runtime rather than the older monolithic conversation loop, which made it easier to build custom orchestration patterns and plug in different model providers. It has 60,472 GitHub stars and 9,116 forks, among the largest communities of any open-source agent framework, and ships under the permissive MIT licence for code (the repo's top-level CC-BY-4.0 licence file covers documentation/content only, not the code itself).

As of this listing, the microsoft/autogen README carries a maintenance-mode banner: "AutoGen is now in maintenance mode. It will not receive new features or enhancements and is community managed going forward. New users should start with Microsoft Agent Framework." This isn't a stale badge — it matches the repo's actual activity. The last pull request merged into microsoft/autogen was #7521 on 2026-04-06; nothing has merged since, despite dozens of open community PRs sitting untouched. The autogen-agentchat and autogen-core packages on PyPI have not been updated since 2025-09-30 (currently v0.7.5). Microsoft's replacement, Microsoft Agent Framework, unifies AutoGen's orchestration ideas with Semantic Kernel's enterprise plumbing, reached general availability on 2026-04-03 with "stable APIs and a commitment to long-term support," and is the framework Microsoft is actively investing in (github.com/microsoft/agent-framework: 12,856 stars, commits as recently as 2026-08-17). Microsoft publishes a migration guide from AutoGen to Agent Framework.

In late 2024, AutoGen's original creators (Chi Wang, Qingyun Wu, and others) split from Microsoft's stewardship over governance — wanting, in Wang's public account, to "move faster and make decisions more efficiently without corporate constraints." They settled on a new home, github.com/ag2ai/ag2, under the Apache-2.0 licence, and there was a public dispute over the pyautogen PyPI package name that Microsoft eventually regained control of (the package now just redirects installers to autogen-agentchat); the bare autogen PyPI name is still held by AG2. AG2 is smaller by community size (4,869 stars, 699 forks) but is where the original API and community are still actively developed: its last push was 2026-08-17, with multiple PRs merged in the days before this listing was written. If you specifically want the classic conversable-agent AutoGen programming model kept alive and evolving, AG2 rather than Microsoft's repo is where that's happening now.

Code is MIT-licensed; documentation/content in the repo is separately licensed CC-BY-4.0 — check which file governs what you're reusing. A standard Microsoft SECURITY.md routes vulnerability reports through MSRC ([email protected]); it's a real disclosure process, not a compliance/trust page (no SOC 2, DPA, or sub-processor list — this is a framework you self-host, not a hosted service with that kind of surface). A search of OSV.dev and GitHub Security Advisories found no published CVEs against pyautogen, autogen-agentchat, autogen-core, or ag2 as of 2026-08-18, though that reflects no indexed advisories rather than a guarantee of none existing.