Microsoft's AutoGen agent framework archived with no fixes as AI Frontiers group ends

Microsoft's AutoGen, one of the most-starred open-source projects in the AI agent ecosystem, has been abandoned. A banner now on the project's README states that the framework is archived and that Microsoft will not fix bugs, patch security issues, add features, or offer support — a quiet ending for a repository with 61,747 GitHub stars and 1,171 open issues.

The details come from a dev.to analysis published November 8, which reconstructed the timeline from the GitHub API, PyPI and a blog post by the group's co-founder.

The research group behind it has ended

John Langford, who helped found the AI Frontiers group inside Microsoft Research in 2023, declared the group finished in a November 6 post on his hunch.net blog. In his account, the roughly 50-person team built AutoGen and the Fara computer-use agent, alongside research efforts including the Full Bandwidth Transformer, the Next Latent objective and Dion3. In early September the group was folded into Microsoft AI, the division led by former Inflection leadership. Langford characterizes that leadership's approach as favoring scaling to the explicit exclusion of research, says the group's other founding leaders came to similar conclusions about leaving, and notes there will be no internships with the group next year.

One caveat matters: the group's end is Langford's account. Microsoft has issued no formal closure statement — but the repository evidence gathered in the dev.to analysis backs up the substance.

The repository told the story before the announcement

The last functional code change to AutoGen landed March 26, 2026. Between April 6 and October 16 there were zero commits of any kind. The only activity in the second half of the year consisted of two documentation changes, one clarifying that the project is archived and one adding an archive caution to the landing page. On PyPI, autogen-agentchat last shipped version 0.7.5 on September 30, 2025 — more than 400 days ago, with nothing since.

The README also contradicts itself. Its banner disclaims all fixes and support, while a contributing section further down still invites bug fixes, security patches and documentation improvements. GitHub's platform-level archive flag, meanwhile, remains unset, meaning the archived status exists only as prose. That nobody has reconciled the two statements is itself a measure of how little attention the repository now receives.

Fara and the successor

Fara, the group's other project, is in better shape: 6,251 stars, MIT-licensed, 44 open issues, and a Fara1.5 release on July 22, 2026 delivering computer-use agent models in 4B, 9B and 27B sizes with weights on Hugging Face. Its most recent push came September 22, 2026, and its README carries no archive notice — development there has slowed rather than stopped.

Microsoft's designated replacement is Agent Framework, which reached version 1.0.0 for Python and .NET on April 2–3, 2026 and has shipped steadily since, with python-1.25.0 released November 7 and dotnet-1.28.0 on November 6. AutoGen's README points existing users to a migration guide. As the dev.to analysis puts it, the handover happened in April; October's archive notice simply made it official.

Why it matters

For teams with AutoGen in production, the banner removes any ambiguity: no security patches are coming, and migrating to Agent Framework — or another orchestration library — is now a planning item rather than a preference.

The episode also says something about Microsoft's direction. Langford's underlying argument is that research can outperform raw scale — he cites ByteDance experiments suggesting a Full Bandwidth Transformer variant could match a model twice its size, halving training and inference cost. Microsoft's revealed preference, per the dev.to analysis, ran the other way: the scaling organization absorbed the research group, its founders left, and the framework the researchers built now redirects users to the scaling organization's replacement.

Finally, it is a reminder that GitHub stars measure adoption, not maintenance. A repository can carry 61,747 stars, a live README and a contributing guide while receiving nothing but archive notices — and the only support its dependents will ever get is an instruction to migrate away.