Microsoft Disbands Its Flagship AI Frontiers Research Group as Co-Founder John Langford Leaves
Microsoft’s flagship AI research group is over. John Langford, the Microsoft Research leader who helped create AI Frontiers, announced on November 6 that he has left the company — and that the group itself is finished. The roughly 50-person team behind the AutoGen agent framework and the Fara computer-use agent was folded into Microsoft AI in early September. Langford says the two groups simply could not work together, because one was built for research and the other for scaling.
Microsoft has not commented publicly on the change. Everything we know comes from Langford’s own account, published on his blog, Machine Learning (Theory). But the facts he lays out are specific, and they carry a warning for anyone building on the tools his team made.
What Langford announced
In a post titled “The End of AI Frontiers”, Langford wrote that the group was set up inside Microsoft Research after ChatGPT launched, to run research in a more coordinated way. In early September, it was reorganised into Microsoft AI, the much larger organisation led by former Inflection executives.
Langford says Microsoft AI’s leadership centres on scaling — bigger models, better training data, longer training runs — and that this made the two groups incompatible. He writes that he placed his team members as best he could and left, and that every other founding AI Frontiers leader reached a similar conclusion about leaving. He also says there will be no AI Frontiers internships in the coming year.
What AI Frontiers built
AI Frontiers was not a side project. Formed in 2023, it worked on agents, core models and shared infrastructure. Its best-known work is AutoGen, the multi-agent orchestration framework that became one of the most popular open-source agent projects, and Fara, a computer-use agent. The team also worked on model research including the Full Bandwidth Transformer, Next Latent and Dion3, plus an edge-scale pretraining pipeline.
That makes the shutdown more than an org-chart story. When a research lab closes, the code it maintained does not vanish — but the people who understood it do. Langford considered returning to Microsoft Research before leaving, and says he did not go back because research there has no clear path to use.
Why scaling won
The tension Langford describes is the central argument in AI right now: does progress come from bigger models and more compute, or from new ideas? Microsoft AI’s leadership has bet on the first answer. Langford’s argument is that research still pays — he points to tests where a Full Bandwidth Transformer variant matched a model twice its size, which would halve training and inference costs.
Whether or not that claim holds up, the direction of travel is clear. The same month that saw Anthropic’s IPO valuation slide and OpenAI’s funding round fall short is now seeing the industry’s biggest companies reorganise around scale. Research labs that cannot show a path to a product are losing their place.
What happens to AutoGen now
This is the practical question for developers. AutoGen was already winding down before Langford’s post: the last release of autogen-agentchat on PyPI was version 0.7.5 in September 2025, and on October 16 a change to the AutoGen repository stated plainly that “Microsoft is not providing bug fixes, security fixes, new features, or support for this project.”
Langford’s post does not change AutoGen’s status, and it does not say what happens to Fara or to the people who maintain these projects. The Fara repository is MIT-licensed and not archived, with its last push in September. But with the research group around them gone, anyone building on these frameworks should plan as if support is finished — because, by Microsoft’s own README, it already is.
What it signals
A vendor’s org chart is part of your AI stack. Agent frameworks can lose their team overnight, and the difference between “just scale” and “make it cheaper” bets is now a real fork in the road for the whole industry. If you build on open-source AI tools, this week is a reminder to stay model-agnostic and to check who is actually maintaining the code you depend on.