Microsoft disbands AI Frontiers research group as founder John Langford departs

Microsoft's AI Frontiers research group is no more, and John Langford, the Microsoft Research leader who helped found it, says he has left the company along with it. According to a report on dev.to, Langford announced his departure on November 6, and the roughly 50-person team behind the AutoGen agent framework and the Fara computer-use agent had already been folded into Microsoft AI in early September. Langford's explanation, as relayed by dev.to, is that the two organisations were built for different jobs — one for research, one for scaling — and could not do those jobs under one roof.

Why the group ended

In a farewell post titled "The End of AI Frontiers" on his blog, Machine Learning (Theory), Langford described how the group came about: after ChatGPT launched, AI Frontiers was set up inside Microsoft Research so that research could be run in a more coordinated way. Microsoft AI, by contrast, is led by former Inflection executives and is oriented around scaling — larger models, better training data and longer training runs. According to dev.to's account, Langford concluded the two missions were incompatible.

He writes that he found placements for his team members as best he could before leaving, that every other founding leader of AI Frontiers came to the same conclusion, and that the group will not run internships next year. Microsoft has not commented publicly, dev.to reports.

What the group leaves behind

AI Frontiers was formed in 2023 and worked on agents, core models and shared infrastructure. Its best-known output is AutoGen, among the most widely used open-source frameworks for multi-agent systems, along with Fara, a computer-use agent. The group also produced model research — the Full Bandwidth Transformer, Next Latent and Dion3 — plus an edge-scale pretraining pipeline.

The dev.to report argues this is what lifts the story beyond an internal reorganisation: when a lab shuts down, the code stays but the expertise walks out the door. Langford reportedly considered returning to Microsoft Research but declined, on the grounds that research there has no clear route into practical use.

The argument behind the split

The friction Langford describes mirrors the central dispute in AI today: does progress come mainly from bigger models and more compute, or from new ideas? Microsoft AI's leadership has backed the first answer, per dev.to. Langford contests it, pointing to tests in which a Full Bandwidth Transformer variant performed on par with a model twice its size — a result that, if it holds generally, would halve training and inference costs.

The practical fallout for AutoGen users

For developers, the immediate question is what happens to AutoGen. According to dev.to, the project had already been winding down before the group dissolved: the final release of autogen-agentchat on PyPI was version 0.7.5 in September 2025, and a change made to the AutoGen repository on October 16 stated plainly that Microsoft is not providing bug fixes, security fixes, new features or support for the project.

Fara's repository, dev.to notes, is MIT-licensed and has not been archived, with its most recent push in September. But with the research group that produced these tools now gone, the report's advice to anyone building on them is to plan as though support has already ended — because by Microsoft's own repository documentation, it has.

Why it matters

The dissolution of AI Frontiers is a data point in a broader fight over how AI progress happens: the scaling-first view that now dominates Microsoft AI won an internal argument, and a research structure created to explore alternatives did not survive contact with it. It also shows how quickly a large vendor can shed a research capability, along with the people who understood its code.

It is equally a maintenance warning for anyone running agents in production. Frameworks can lose the team behind them overnight, and a vendor's org chart is effectively part of your stack. The dev.to report's closing recommendation is sound general practice: stay model-agnostic where possible, and check who is actually maintaining the code you depend on.