0 fixes for 61,747 stars: AutoGen after AI Frontiers died

AutoGen has 61,747 stars and 1,171 open issues. Its README now says Microsoft provides no bug fixes, no security fixes, no new features, and no support. Its last PyPI release is 404 days old. The research group that built it is over — and I checked the repos to see what "over" actually means, because the group died twice, and the two deaths are six months apart.

The announcement

John Langford, who helped create AI Frontiers inside Microsoft Research in 2023, published "The End of AI Frontiers" on his hunch.net blog on November 6. The facts in it are plain: the group was about 50 people; most of its work was figuring out how to make agents work — the AutoGen orchestration framework and the Fara computer-use agent — plus fundamental work like the Full Bandwidth Transformer, the Next Latent objective, and Dion3. In early September it was reorganised into Microsoft AI. Langford calls the ex-Inflection leadership's philosophy "scaling, and aggressively not research", says every other founding leader reached a similar conclusion about leaving, and writes that there will be no internships with the group next year.

One distinction matters: Langford declared the group over. Microsoft has published no formal closure announcement. That is his account, and I am treating it as such — but the repos back up the substance of it.

The repos died first

Here is AutoGen's commit history, from the GitHub API, most recent first:

2026-10-28 d07c667 docs: add archive caution to landing page (#8393) 2026-10-16 004749e docs: clarify AutoGen is archived and no longer maintained (#8366) 2026-04-06 027ecf0 Update maintenance mode banner in readme (#7521) 2026-03-26 8544314 fix: restrict importlib provider loading to trusted namespaces (#7463) 2026-03-11 b047730 fix: Improve AutoGen Studio: deprecate FunctionTool, harden MCP WebSocket endpoint 2025-10-04 13e144e fix: order by clause (#7051) 

Read that list again. The last code change landed March 26 — 227 days ago. Between April 6 and October 16 there were zero commits of any kind. The only activity in the second half of 2026 is two documentation commits that say the project is archived. And the PyPI package tells the same story: autogen-agentchat 0.7.5, uploaded September 30, 2025 — 404 days ago, with nothing since.

So the timeline is: the code stopped in March, the releases stopped in September 2025, the archive banner went up October 16, and the founder said the group is over on November 6. The announcement is the last event in the sequence, not the first.

The README contradicts itself

The current README, fetched live today, opens with an "Archived" badge and a caution block: "This project has been archived and is no longer actively maintained by Microsoft. Microsoft is not providing bug fixes, security fixes, new features, or support for this project." But the GitHub API still reports archived: false — the repo carries the word "archived" as prose, not as a platform state. And further down the same README, the contributing section still reads: "As AutoGen is in maintenance mode, contributions are limited to bug fixes, security patches, and documentation improvements."

Both cannot be true. Either the project takes bug fixes or it does not; the banner says it does not, a stale paragraph says it does, and 1,171 open issues sit between the two claims. If you depend on AutoGen, the operative sentence is the banner, not the paragraph — but the fact that nobody has cleaned up the contradiction tells you how much attention the repo gets now. None.

What about Fara?

Fara, the other named product of the group, is in a different state: 6,251 stars, MIT-licensed, created October 29, 2025, last push September 22, 2026, 44 open issues. Its README leads with Fara1.5, released July 22, 2026 — a family of computer-use agent models in 4B, 9B, and 27B sizes, with weights on Hugging Face and two benchmark datasets. No archive notice anywhere in it. Fara is quiet, not dead: a model family with a July release and a September push is a finished artifact more than an abandoned one. The group is over; the models are still downloadable.

The successor is shipping weekly

Microsoft's answer to "what replaces AutoGen" is Agent Framework, which hit 1.0.0 for Python and .NET on April 2–3, 2026. Its release cadence is the mirror image of AutoGen's:

AutoGen: last code commit Mar 26, 2026; last PyPI release Sep 30, 2025 Agent Framework: python-1.25.0 released Nov 7, 2026 (yesterday) dotnet-1.28.0 released Nov 6, 2026 1.0.0 shipped Apr 2-3, 2026 

The README points existing users at a migration guide. The handover happened in April; the archive notice in October just made it official.

The argument underneath

Langford's post is not really about a reorg. It is about a bet: "returns to scale are far more predictable than returns to research." He answers with the ByteDance experiments suggesting a Full Bandwidth Transformer variant might match a model twice its size — halve the training and inference cost, on an industry investing something approaching a trillion dollars. Do the math, he says. He is right that the math is the argument: if research halves cost, it is worth more than another scale-up. And Microsoft's revealed preference is the opposite: the scaling organisation absorbed the research group, the research group's founders left, and the framework the researchers built now tells you to go use the scaling organisation's framework instead.

That is what "over" means here. Not that the code vanished — 61,747 stars do not vanish, and the repo is still readable. It means the institution that would have answered the next question about agents is gone, the people who would have answered it work elsewhere, and the only maintenance AutoGen will ever get is a banner telling you to migrate.

Sources: John Langford, "The End of AI Frontiers", hunch.net (Nov 6, 2026, read in full); microsoft/autogen — GitHub API (stars, forks, archived flag, commits, open issues) and README fetched Nov 8, 2026; PyPI autogen-agentchat release history; microsoft/Fara — GitHub API and README; microsoft/agent-framework releases. All counts and dates above are my own reads of those sources; the claim that the group is over is Langford's, and Microsoft has not published a formal closure statement.