Microsoft's AI Frontiers research lab is over, a founding leader says

Microsoft Research’s AI Frontiers lab, the roughly 50-person group behind the AutoGen agent framework and the Fara computer-use models, is over, according to a blog post by one of its founding leaders. The post says the group was reorganised into Microsoft AI (MAI) in early September, that the founding leadership has since left, and that “AI Frontiers is over.”

Microsoft’s own lab page confirms the first half of that story: an update dated Sept. 14, 2026 says AI Frontiers “is now part of Microsoft AI” and that it would pursue the same research goals there. None of the sources include a Microsoft statement about the group ending, so the closure itself rests on the founding leader’s first-person account.

What the founding leader says happened to AI Frontiers

The post, published on the long-running machine learning blog hunch.net, describes a sequence of events rather than a single announcement. The author writes that the group was moved into Microsoft AI in early September, and that while combining two AI organisations “could make sense” on the surface, in practice “the philosophy of the ex-Inflection leadership of MAI is incompatible.”

“The core philosophy there is scaling, and aggressively not research.”

That is how the author characterises MAI’s approach. In a scaling-only environment, the post argues, most effort goes into “getting/making/refining data,” which the author calls “quite a step away from what we were working on.” The author also writes that “there was plenty of evidence that they just did not really want us.”

The author says they “situated my folks as best I could and left,” and that “every other founding AIF leadership team member seems to have come to a similar conclusion.” The post adds that there will be no internships with the group’s researchers in the coming year.

The account is one person’s view of an internal reorganisation. It does not name the other founding leaders, give a headcount for who left or stayed, or quote Microsoft.

What AI Frontiers worked on

The lab’s X account says it started in October 2023 on one bet: “the interesting question isn’t how big a model gets; it’s what computing looks like when agents and people work together.” The founding leader describes the group’s origin the same way from the inside: after the ChatGPT release in 2023, several researchers concluded that work inside Microsoft Research needed to be more organised to compete, and formed AI Frontiers as a sub-organisation of Microsoft Research.

Microsoft’s lab page says the group focused on “efficient language models and agentic systems” and lists what it shipped:

• AgentInstruct
• The Phi family of models, Microsoft’s small language models
• The Fara portfolio of computer-use agent (CUA) models, which operate software through its interface the way a person would
• AutoGen, an agent framework for building systems of multiple cooperating AI agents
• Magentic-One, MagenticUI and Magentic Marketplace
• Work on agent red teaming, benchmarking and safety research

The page names three research areas: adaptive agents that learn across tasks and modalities, agentic robustness, and “agentic ecosystems” where autonomous agents work alongside people. Saleema Amershi, listed as a Partner Research Manager at Microsoft Research AI Frontiers in a December 2025 Microsoft Research feature, described that last theme as a coming era “where autonomous agents collaborate, negotiate, and transact on behalf of people and organizations.”

The founding leader writes that much of the group’s work was “figuring out how to make agents work,” citing AutoGen and the Fara agent as examples. A team in New York pursued more fundamental model research, according to the post: a “Next Latent” training objective aimed at extrapolation, Dion3, described as an optimised version of the Muon optimiser, and a “Full Bandwidth Transformer” that the author says makes models substantially more capable through recursion.

Recent releases builders may already use

The group was shipping right up to the reorganisation. Posts on its X account describe several releases over the summer:

• MagenticLite open weights (July 22 post): MagenticBrain, a 14B-parameter orchestrator model, and the Fara 1.5 computer-use family in 4B, 9B and 27B sizes, previously on Microsoft Foundry, were published on Hugging Face with open weights. The post says the app, the harness and every model in the stack are now openly available.
• Echoverse (July 31 post): “deep, evolving environments for computer-use agents,” with code on GitHub. The post says a 9B model trained on twelve worlds went from 36.5% to 67.1%; the visible text is cut off before it names the evaluation.
• Web Skill Factory (Aug 4 post): a pipeline that turns solved web tasks into reusable, verified code, so agent “skills” are programs rather than notes a model rereads each run. The post reports results on a WebArena subset with gpt-5.4, but the figure is cut off in the visible text.

These figures are the lab’s own, from its own posts, and none of the sources include an independent evaluation.

AI Frontiers at a glance	Detail
Started	October 2023
Size	About 50 people, per a founding leader
Parent organisation	Microsoft Research, then Microsoft AI from early September
Lab page update	Sept. 14, 2026: “now part of Microsoft AI”
Best-known work	AutoGen, Phi models, Fara CUA models, Magentic-One
Latest open release	MagenticBrain 14B and Fara 1.5 (4B/9B/27B) on Hugging Face

The scaling versus research argument behind the split

The founding leader’s core complaint is about strategy, not just organisation. The post concedes that “there is plenty of evidence that bigger models are better models” and that, from a business view, “returns to scale are far more predictable than returns to research.” But it calls a scaling-only focus “both misdirected and not fun,” and argues that “returns to research are real and significant.”

As evidence, the author points to experiments by ByteDance that, according to the post, suggest a variant of the Full Bandwidth Transformer “might be comparable to a model of twice the size,” implying lower training and inference costs for the same capability. That is the author’s summary of someone else’s experiments, and the sources include no further detail on them.

Microsoft’s framing of the move is different. Its lab page calls MAI “a natural fit for our work, given MAI’s focus on humanist superintelligence and building AI systems focused on real problems and in service to people.” The lab’s own Sept. 2 post greeted the move with “AI Frontiers is now part of Microsoft AI!”

For context on MAI’s priorities, Microsoft AI chief Mustafa Suleyman told CNBC in 2025 that instead of being “the absolute frontier,” Microsoft preferred AI models that are “three to six months behind,” as reported by Pivot to AI.

The post also raises a concern about Microsoft Research as a whole. The author says they considered returning to Microsoft Research but saw two obstacles: no clear “path to real use” for research, and what the author calls “a breakdown of the basic deal in MSR: do good work and good things happen.” The author writes that they “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.”

This is not the first time a specialised Microsoft AI team has been dissolved during a strategic push. In 2023, Microsoft cut its AI ethics and society team as part of 10,000 job cuts, according to Platformer reporting cited by Ars Technica, VKTR and The Register. Microsoft said at the time that it remained committed to responsible AI and that some affected staff moved to other teams. The two cases differ in purpose and size, but both involve a small research-style group losing its standalone status.

What the closure means for builders using AutoGen and Fara

The practical question for developers is what happens to the tools. The sources do not answer it directly.

• Released weights and code stay public. According to the lab, MagenticBrain and the Fara 1.5 models are on Hugging Face with open weights, and the MagenticLite app and harness were released openly. Echoverse code is on GitHub. Published open weights cannot be withdrawn from people who already have them.
• Maintenance is an open question. No source says who, if anyone, now owns AutoGen, Fara or the Magentic projects inside Microsoft AI. The lab’s pinned X post from April 27 addresses the AutoGen community with “an important update,” but the visible text is truncated before saying what it is.
• Shared infrastructure. The founding leader says AI Frontiers built a monorepo shared by default with all of Microsoft Research, including what the author calls a “proven-capable edge scale pretraining codebase.” The post does not say what happens to it.

Teams depending on these projects in production should watch the official repositories for maintainer changes and plan for the possibility of community-led forks rather than assume continued Microsoft investment.

Where the staff and projects go

On staff, the sources say little beyond the founding leader’s account. The author says founding leadership has left, that they placed their own team members “as best I could,” and offers “best wishes for the folks I leave behind,” which suggests some researchers remain inside Microsoft. The author also describes having had trouble getting “innovation into use (or even seriously considered)” while in MAI.

Microsoft’s lab page, written at the time of the move, said “We’ll share more exciting news on this transition soon!” The sources contain no follow-up announcement.

What to watch next

• A Microsoft statement. Whether Microsoft confirms the founding leader’s account, or says which AI Frontiers projects continue under Microsoft AI.
• Repository activity. Commits, maintainer lists and issue responses on AutoGen, Fara, Magentic and Echoverse will show whether the projects stay active.
• Where the founders land. The author writes that “many other people in the world” value research and innovation, implying new roles elsewhere.
• Microsoft Research’s direction. The founding leader’s warning that similar reorganisations may continue at Microsoft Research is a claim worth testing against what the company does next.