One founder's account, stated as such
Langford's post is a departure note, not a company announcement, and it should be read that way. Microsoft is not quoted in it. What he describes is specific enough to check against the record, and the parts that matter most for readers are the ones he states plainly.
AI Frontiers was set up, in his telling, because research inside Microsoft Research was too loosely organized to compete after ChatGPT: a good idea pursued by one person would "arrive late and unconvincing". The group ran coherent efforts from what he calls the messy middle, neither top-down nor bottom-up, with about 50 people and compute he describes as an often-severe bottleneck. Much of its work was, in his words, "figuring out how to make agents work", which produced AutoGen and Fara. A New York team worked on fundamentals: the Next Latent objective, the Dion3 optimizer and the Full Bandwidth Transformer.
The break, in his account, came with the September move. Microsoft AI's leadership, which came from Inflection, holds a philosophy he summarizes as "scaling, and aggressively not research". Returns to scale are predictable; returns to research are not, and he accepts that logic before rejecting it. As an example of research he believes deserves investment, Langford cites experiments at ByteDance suggesting that a Full Bandwidth Transformer variant might achieve performance comparable to a model twice its size. The post does not establish that the result generalizes across workloads or that Microsoft could realize the implied economic benefit. He says he could not get innovation used, or even seriously considered, inside Microsoft AI, placed his people as best he could, and left.
That is an argument about how a large company should spend its research budget, made by someone who lost it. It may be right. It is not the part of this story a working developer needs to act on.
The README already told AutoGen users
AutoGen's repository carries the actionable news, and it predates Langford's post. Under a caution heading, the README states that the project "has been archived and is no longer actively maintained by Microsoft", that Microsoft "is not providing bug fixes, security fixes, new features, or support for this project", and that vulnerabilities in the project and its dependencies "may remain unaddressed". It adds that AutoGen "was developed for research and experimentation" and advises against production use. New users are sent to Microsoft Agent Framework, described as the enterprise-ready successor, and existing users to a migration guide.
That wording arrived in a README change merged on 16 October 2026, in pull request #8366, three weeks before Langford wrote anything, and the last release of the main Python package, autogen-agentchat 0.7.5, was uploaded to PyPI on 30 September 2025. Microsoft's public warning therefore predates Langford's November 6 departure post by three weeks. That establishes when the notice appeared in the README, not the precise date on which maintenance or security support ended. Langford's account provides organizational context for AutoGen's history, but it does not establish why Microsoft ended support or whether the company could reconsider that decision. The repository's own warning is sufficient grounds for developers to reassess their reliance on the framework.
Two details in the same repository cut against a clean reading, and both were visible in a first-hand read on 7 November. The repository metadata still lists the project as not archived, so a tool that checks the archive flag rather than the README text will report AutoGen as live. And a line near the bottom of the README still describes AutoGen as being "in maintenance mode", with contributions limited to bug fixes and security patches, which is exactly what the caution block at the top says Microsoft is not providing. The README's warning is explicit, but the unset GitHub archive flag and the older maintenance-mode wording create conflicting signals for readers and automated inventory tools. The repository has not been marked archived through GitHub's archive mechanism, despite the README describing the project as archived.
AutoGen's naming history creates another source of confusion. The secondary analysis identifies the plain autogen distribution on PyPI with the separate AG2 project, while Microsoft's newer framework packages include autogen-agentchat and autogen-ext. Developers should check package ownership, installed versions and actual dependencies before deciding whether Microsoft's archive notice applies to their deployment. The notice should not automatically be attributed to AG2 merely because the projects share naming history.
What this does and does not settle
Nothing in Langford's post, and nothing in the README, says what happens to Fara, and Microsoft has not, in the sources read for this piece, published its own account of the group's end or its staffing. The staffing claims, the philosophy dispute and the prediction that Microsoft Research will keep reorganizing groups away are Langford's, and the draft of record treats them that way.
What is settled is narrower and more useful. Microsoft's AutoGen README explicitly states that the project is no longer actively maintained and will receive no bug fixes, security fixes, new features or support from Microsoft. That warning applies even though the repository's GitHub archive flag was still unset when the desk checked it on November 7. That fact was published on 16 October in a README existing users may not revisit. If AutoGen is in your stack, the migration guide linked from that README is now the document that matters, and the time to read it is before a dependency scan, or an incident, reads the archive notice for you.
Sources
- [1] John Langford, "The End of AI Frontiers", Machine Learning (Theory), 6 November 2026, read in full for this article: the group's 2023 creation inside Microsoft Research, its roughly 50 people, AutoGen and Fara, the September move into Microsoft AI and his departure. Every organizational claim in this article is his account; Microsoft is not quoted in the post.Read source
- [2] AutoGen repository README and repository details, Microsoft, read in full on 7 November 2026 and re-checked on 8 November: the caution block ("Project archived and no longer maintained"; "Microsoft is not providing bug fixes, security fixes, new features, or support for this project"), the unset GitHub archive flag ("Archived: No"), the leftover maintenance-mode line, 61,743 stars and 9,381 forks.Read source
- [3] Pull request #8366, "docs: clarify AutoGen is archived and no longer maintained", microsoft/autogen, merged 16 October 2026; merge date verified in the repository's own pull-request record for this article.Read source
- [4] autogen-agentchat on PyPI: latest release 0.7.5, uploaded 30 September 2025 (PyPI record checked for this article). The plain autogen package's own PyPI page describes the separate AG2 project, not Microsoft's framework.Read source
- [5] Gabriel Anhaia, "The Microsoft Lab That Worked on AutoGen Is Over. AutoGen Already Gets No Security Fixes", dev.to, 7 November 2026: the secondary analysis that first connected Langford's post to the README change and identified the plain autogen package with AG2.Read source