Research group behind Autogen and Fara dissolves inside Microsoft

From creation to end, in three years

In 2023, shortly after the launch of ChatGPT, a group of Microsoft Research researchers concluded that the field needed more organization to compete on speed with other AI labs. That is how AI Frontiers (AIF) was born, a subdivision of about 50 people dedicated to coordinated research projects, according to an account published on the blog hunch.net, by John Langford, a researcher who helped found the group.

In early September 2026, AIF was reorganized and absorbed into Microsoft AI (MAI), the company's division led by former Inflection AI executives. In Langford's view, merging the two teams made sense on paper: each was ahead of the other in different areas. In practice, he says, the philosophy of the leadership coming from Inflection is incompatible with the work AIF was doing.

In the weeks following the merger, Langford left Microsoft. According to the post, all the other members of AIF's original leadership reached the same conclusion and also left the company. The group will no longer offer the internship program it used to run every year.

What the team had on its hands

AIF was not a small team of pure theory: it built tools and research that already circulate outside Microsoft. Among the projects mentioned in the post:

• Autogen: an orchestration framework for applications with multiple AI agents, now used by developers outside Microsoft to build agentic systems.
• Fara: a computer use agent, capable of operating graphical interfaces on its own.
• Next Latent objective: a training objective aimed at extrapolating model capabilities.
• Dion3: an optimized version of the Muon algorithm, used in model training.
• Full Bandwidth Transformer: an architecture that uses recursion to make models more powerful without increasing their size.

The group also created a monorepo shared by default with all of Microsoft Research, including a large-scale pretraining pipeline that Langford presented in October, in the keynote at COLM (Conference on Language Modeling). According to him, there are at least four efforts in total inside Microsoft trying to build equivalent pretraining pipelines, including AIF's own pipeline (which is public within MSR); the other three are not public.

Research versus scale: the fracture with Microsoft AI

The central friction, according to Langford, is philosophical. The leadership coming from Inflection AI prioritizes scaling models (more data, more parameters) and treats research as secondary. For him, this is an understandable but mistaken reading of the trade-off between innovation and size.

I consider the scaling-only focus to be both misdirected and not fun.

John Langford, former leader of AI Frontiers at Microsoft Research

As a counterpoint, Langford cites a concrete example: experiments conducted by ByteDance suggest that a variant of the Full Bandwidth Transformer achieves quality comparable to a model with twice its size. If confirmed at production scale, this would mean cutting training and inference cost in half for the same capacity, a relevant figure for an industry that already invests close to a trillion dollars in AI infrastructure.

In short: on one side, a bet on research that promises efficiency gains that are hard to capture with more GPUs alone; on the other, a bet on scale that is more predictable from a business standpoint, even if it costs more per unit of capacity.

What remains open for those using Autogen and Fara

Here is the point that matters to anyone who has already built something on top of these tools: Autogen is an open-source project used by developers to orchestrate multiple AI agents in real applications, and Fara was presented as an agent capable of operating graphical interfaces. With the departure of all the leadership that created and maintained these projects, Langford's post does not provide details on what happens, going forward, to the maintenance or roadmap of these frameworks inside Microsoft.

For teams in Brazil that already have projects running on Autogen, this does not mean the code will stop working tomorrow: open-source repositories remain accessible regardless of the company's internal structure. But it is worth watching the pace of commits and releases over the coming months as a practical signal of how much official support the project will still receive.

Langford closes the account with a personal note: it is not hard to imagine a future Microsoft that becomes attractive for research again, should the leadership develop, in his words, a bit more wisdom about how to organize teams for the industry's current moment.