Microsoft ends AI Frontiers: no more room for research at Microsoft AI

Microsoft Research has lost one of its most interesting teams: AI Frontiers, a group of roughly 50 researchers working on agents and new ways of training models, is officially over. On November 6, 2026 its founder John Langford — a Microsoft researcher and the author of Vowpal Wabbit — announced the end, and the entire founding leadership followed him out the door. The reason is sobering: in September the team was non-consensually reorganized into Microsoft AI, whose Inflection-derived leadership bets only on scaling models and, in Langford’s words, is “aggressively not research”.

Why AI Frontiers existed

The prologue is 2023, right after ChatGPT launched. Langford noticed something uncomfortable: lone research efforts inside Microsoft Research rarely survived. A single person with a clever idea arrives late and unconvincingly — killing the idea is easier than testing it, because the baselines one person can assemble are too weak. Together with colleagues he set up AI Frontiers inside MSR: neither “top down” nor “bottom up”, but in the “messy middle”, where efforts are deliberately organized into coherent directions instead of being scattered across individual projects.

Compute was a recurring bottleneck in the ~50-person group, but by his own estimate it managed to do pioneering things that MSR could not have carried to a convincing result in its usual form.

What the group shipped: from AutoGen to the Full Bandwidth Transformer

Most of AI Frontiers’ work was about agents. The AutoGen orchestration framework, on which many Microsoft multi-agent applications are built, grew out of that work, as did Fara, an agent that operates a computer directly. Alongside it, the NYC part of the group pushed fundamental directions: the Next Latent objective, which lets models extrapolate beyond what they saw; Dion3, an improved and optimized version of the Muon optimizer; and the Full Bandwidth Transformer, where a model performs substantially more powerfully through recursion.

The group deliberately shared with all of MSR: it kept a monorepo open by default to every researcher in the company. Inside it sat a proven-capable edge-scale pretraining pipeline, which Langford showed in the COLM keynote last month. Building such a pipeline properly, he notes, takes enormous effort — and at least four separate teams inside the company were trying to build one independently.

How it ended

In early September 2026 AI Frontiers was reorganized into Microsoft AI. On paper the logic holds: merge two AI organizations, concentrate effort, exploit MAI’s far larger resources. In practice, two incompatible philosophies collided. MAI’s Inflection-derived leadership believes better models come primarily from scaling — more data, more training, predictable returns — and from making and refining that data. Research does not fit that picture, because returns to research are less predictable than returns to scale. Langford pushes back with facts: ByteDance experiments suggest a variant of the Full Bandwidth Transformer can match a model twice its size, halving training and inference cost at the same capability. “Do the math — how much is it worth to halve training and inference cost, given that industry investment is approaching a trillion dollars?” he writes.

Events then moved fast. Langford says there was “plenty of evidence that they just did not really want us”, and that he placed his people as best he could and left. Every other founding member of the AIF leadership team apparently reached the same conclusion.

Why they did not return to MSR

The most interesting question is why the experiment could not simply be rebuilt. Langford gives two reasons. First: at a company where his research has no path to real use — and is not even seriously considered — there is nothing to return to. Second: the informal deal inside Microsoft Research, “do good work and good things happen”, is broken. The work was good, and the team was moved away non-consensually anyway. “I would not be surprised to see that kind of thing continue to happen until MSR leadership develops a will to effectively organize to meet the moment,” he writes. AI Frontiers internships are cancelled; the group, in his words, is over.

Why it matters beyond Microsoft

The story is less about 50 people than about what regularly happens to research groups when they are folded into product structures that live on scaling. The “innovation versus size” trade-off is solved, in the moment, in favor of size — an understandable, predictable failure mode. But returns to research are real and significant, and every exit of this kind subtracts from the probability of breakthroughs that later get credited to scale anyway.

Tellingly, the journey did not end in a shuttered lab after a failure, but in the absorption of a successful group and the departure of all of its founders. For agents — one of the hottest parts of the market — Microsoft just lost the team that built agent orchestration and computer control. Langford’s original post is on his Machine Learning (Theory) blog (November 6–7, 2026); the departure and the group’s end have been confirmed by several independent outlets.

What this means for users

Access to the group’s internal work remains closed, but the agentic models against which this story plays out are available right now in NeuralSpace Chat. Testing a modern model on your own task is, by the way, the most honest way to feel “returns to research” for yourself.

And a bit of real statistics. According to anonymized data from NeuralSpace’s production database over the last 30 days (window 10.10–09.11.2026), the platform handled 17,215 image generations, of which 16,336 (94.9%) completed successfully. The median wait for a finished image was 62 seconds, with a 95th percentile of 217 seconds. Video generation accounted for 2,860 tasks in the same period, 2,207 of them (77.2%) completed successfully, with a median of 81 seconds and a 95th percentile of 495 seconds.

FAQ

What was AI Frontiers?

A sub-organization inside Microsoft Research that John Langford put together in 2023 to turn scattered research initiatives into coherent efforts with about 50 people. It produced AutoGen and Fara and worked on new training objectives and optimizers.

Why did the group end?

In September 2026 it was reorganized into Microsoft AI, whose leadership believes model quality comes from scaling rather than research. The founders concluded their direction would not survive under those conditions and left.

What happens to AutoGen and Fara?

Both are already open or productized projects and are not directly dependent on the group’s closure. But agentic research inside Microsoft now sits with MAI, which has different priorities.

Who is John Langford?

A Microsoft Research scientist, a key figure in machine learning, the author of the vowpal_wabbit library and the founder of AI Frontiers. He wrote the post announcing the end of the group.