John Langford, one of the group's organizers, published "The End of AI Frontiers" on his blog on November 6, 2026, and updated it the next day. He writes that he and others formed the AI Frontiers sub-organization within Microsoft Research in 2023, because he had come to think research needed to be more organized to succeed than was typical at MSR. The group was small, at about 50 people, and worked on "figuring out how to make agents work, for example the Autogen orchestration framework and the Fara computer use agent."
What he says happened
Langford writes that in early September the group was reorganized into Microsoft AI, led by people from Inflection. He says this could have made sense on paper, then describes the result in blunt terms: "The core philosophy there is scaling, and aggressively not research." He adds that "there was plenty of evidence that they just did not really want us." He says he placed his people as best he could and left, and that "Every other founding AIF leadership team member seems to have come to a similar conclusion."
His closing is short: "So AI Frontiers is over." He says there will be no internships with the group in the coming year.
The other side of the record
Microsoft's public description is very different. Its lab page, updated September 14, says AI Frontiers "is now part of Microsoft AI" and that the team will "pursue our same goal of frontier and cutting edge agentic and model research". The page calls MAI "a natural fit for our work". We found no Microsoft statement responding to the November post, and the page may simply not have been updated.
The argument under the org chart
Langford's dispute goes beyond one reorganization. He concedes the business logic, saying that "returns to scale are far more predictable than returns to research", and he then argues that research still pays. His example is a claim that ByteDance experiments suggest a variant of the Full Bandwidth Transformer "might be comparable to a model of twice the size". That is his summary of someone else's experiment, and we have not verified it. If something like it holds, halving training and inference cost for the same capability would be worth a great deal at current spending.
The same tension runs through the sector. METR's November 5 review of Anthropic found no acceleration in the scale of training compute behind its estimated speedup, which is at least compatible with the view that methods still matter.
What remains uncertain. This is a single first-person account by a departing leader, and his characterisation of Microsoft AI's philosophy is his opinion. We do not know how many of the roughly 50 staff remain, what happens to the shared training code he mentions, or whether Microsoft will reorganize agent research around another group.
Analysis: For teams depending on AutoGen-era research, check who now maintains the projects you rely on, rather than assuming a lab's name guarantees continuity. Leadership teams can take a narrower lesson. A shared monorepo, which Langford says AI Frontiers built for all of MSR, is an asset that a reorganization can quietly strand, so its ownership should be written down before the org chart changes.
Sources and evidence
John Langford wrote on November 6 that Microsoft Research's roughly 50-person AI Frontiers group, behind AutoGen and Fara, is over after being reorganized into Microsoft AI in September, and that he left. His claim of a scaling-only culture is opinion. Microsoft's September page calls the move a natural fit and promises the same research goal. We found no Microsoft response.