Microsoft's AI Frontiers Is Over — and AutoGen Users Just Lost Their Safety Net
Microsoft's AI Frontiers research group is over. That is the verdict of John Langford, the researcher who helped create the group, in a blog post published on 6 November 2026 titled "The End of AI Frontiers". The 50-person group, formed inside Microsoft Research in 2023 after the release of ChatGPT, was moved into Microsoft AI in early September 2026. Langford says he placed his team as best he could, and then he left. Microsoft has not commented, so this is Langford's account, not a company announcement.
What Langford actually said
Langford's post is short, personal, and unusually blunt. He writes that in 2023, after ChatGPT arrived, it was clear that research needed to be more organised to succeed than was typical inside Microsoft Research: a single researcher could not move fast enough or assemble strong enough baselines to prove an idea worth pursuing. Several colleagues agreed, and together they organised AI Frontiers as a sub-organisation inside Microsoft Research.
The group tried to run coherent research efforts, and Langford says the work was succeeding, particularly given its small size. Then, in early September 2026, it was reorganised into Microsoft AI. Langford describes the leadership there, drawn largely from the former Inflection team, as focused on scaling, and aggressively not on research. The two philosophies were incompatible, he writes, and the new leadership did not really want the group. On X, he called the situation "a bit of a train wreck".
He also explains why he did not simply return to Microsoft Research: he could not get innovation into use while inside Microsoft AI, and he fears the basic deal there has broken down. "As far as I can tell we did good work and were non-consensually reorged away," he writes, adding that every other founding member of the leadership team seems to have come to a similar conclusion. His final line is the one that will be quoted most: "So AI Frontiers is over. I'm sorry to say that there will not be internships with the folks there in the coming year."
What AI Frontiers actually built
AutoGen is Microsoft's open-source framework for building systems in which several AI agents work together, passing messages, calling tools, and dividing up a task. It was one of AI Frontiers' best-known products, and it became widely used by developers building multi-agent applications. The group's other headline project was Fara, a computer-use agent that can operate a computer interface on a user's behalf.
The group also did fundamental research, much of it from its New York team: the Next Latent objective, which enables a model to extrapolate; Dion3, an improved version of the Muon optimiser; and the Full Bandwidth Transformer, which he describes as making models function substantially more powerfully through recursion. He argues this research had real economic value: experiments from ByteDance suggest a variant of the Full Bandwidth Transformer might match a model twice its size, saving large amounts on training and inference. His point is that returns to research are real and significant, even when returns to scaling are more predictable.
 |
| Empty research office desk at dusk, laptop showing code, whiteboard of agent diagrams |
AutoGen was already on its own
For developers, the most important part of this story happened three weeks before Langford's post. On 16 October 2026, a change to the AutoGen repository's README stated plainly: "Microsoft is not providing bug fixes, security fixes, new features, or support for this project." The same change described AutoGen as archived and no longer maintained by Microsoft, and added that it was never intended for production use. The README now warns that vulnerabilities may go unaddressed and advises against production use.
That change ended a wind-down running for more than a year. The last release of the autogen-agentchat package on PyPI was version 0.7.5, uploaded on 30 September 2025. The README now points new users to the Microsoft Agent Framework, which it calls the enterprise-ready successor to AutoGen. Microsoft shipped Agent Framework 1.0 on 3 April 2026 for .NET and Python, developed by the core AutoGen and Semantic Kernel teams at Microsoft.
One detail could mislead a dependency audit: GitHub does not treat the AutoGen repository as archived, so there is no read-only banner, and a tool that only checks the archive flag will miss the warning. The Fara repository is MIT-licensed and not archived, with its last push on 22 September 2026. Langford's post says nothing about what happens to either project, or to the people who maintain them.
| Date | What happened to AutoGen |
|---|---|
| 30 September 2025 | Last autogen-agentchat release, version 0.7.5, on PyPI |
| 3 April 2026 | Microsoft Agent Framework 1.0 ships for .NET and Python |
| 16 October 2026 | AutoGen README states no bug fixes, security fixes, new features, or support |
| 6 November 2026 | Langford announces AI Frontiers is over and that he has left Microsoft |
Research versus scaling at Microsoft
The disagreement at the heart of Langford's post is a genuine argument about how a company the size of Microsoft should improve its AI. The scaling view says bigger models are better models, and that the returns to scale are far more predictable than the returns to research. From a business perspective, that is a strong argument: predictable returns are easier to plan, fund, and sell.
The research view, which Langford defends, says new ideas can change the economics entirely: a better architecture can be worth far more than another round of scaling. His complaint is not that scaling is wrong, but that treating it as the only priority is misdirected. That tension echoes the wider argument over whether the AI boom is sustainable: our earlier analysis asked directly whether the is the AI bubble bursting. What makes Langford's post unusual is that it comes from inside one of the biggest AI spenders in the world, from the researcher who built the group he is now declaring over.
 |
| Developer's hands typing on a laptop showing a multi-agent framework with connected nodes |
What this means if you run AutoGen today
The practical message is simple: do not treat AutoGen as a supported framework. If your service runs it in production, the README warning is explicit, and security vulnerabilities may go unaddressed. First, find out exactly which AutoGen you run, because the name covers more than one project. Microsoft's packages use a hyphen, such as autogen-agentchat and autogen-ext. The plain autogen package on PyPI is a different project, AG2, and is not affected. If your environment shows packages starting with autogen-, or pyautogen, the README warning applies to you.
The supported path forward is the Microsoft Agent Framework, whose migration guide documents the move side by side with AutoGen code. The basics carry over, but four defaults differ: the new Agent is multi-turn by default, it is stateless between calls so history must be carried in an AgentSession, teams become a typed Workflow graph, and distributed execution is still planned. A port is mechanical in places, but the changed defaults can alter how many model calls a task makes, so review it carefully.
The wider pattern of departures
Langford's exit is not isolated: NewsBytes notes that Sarah Bond and Eric Boyd have also left Microsoft in recent months. The stakes are high: the Microsoft vs Google vs Amazon cloud earnings show how closely the giants compete for enterprise customers, while OpenAI's $11 billion first close valued it at 1.25 trillion dollars and Anthropic files for its IPO as the market watches whether funding turns into durable revenue.
 |
| Technician carrying a box down a data-centre corridor between server racks |
Key Takeaways
- John Langford announced on 6 November 2026 that Microsoft's AI Frontiers research group is over, after it moved from Microsoft Research into Microsoft AI in early September.
- AI Frontiers had around 50 people and built AutoGen, the multi-agent framework, and Fara, a computer-use agent, plus research including the Full Bandwidth Transformer.
- Microsoft has not commented publicly; this is Langford's own account, not a company announcement of a shutdown.
- AutoGen already receives no bug fixes, security fixes, new features, or support, per a README change on 16 October 2026, which advises against production use.
- Microsoft points developers to the Agent Framework, which shipped version 1.0 on 3 April 2026, but a port needs care because multi-turn behaviour, session state, and workflows differ.
- Langford's core argument is that research and scaling are both needed, and that a scaling-only strategy wastes work that can halve training and inference costs.
Conclusion
AI Frontiers was a small group with an outsized footprint: a framework thousands of developers rely on, a computer-use agent, and research its founder says could make models far cheaper to train and run. Its end, as described by the man who built it, was not a technical failure but an organisational one. For developers, the test is immediate: check what you run, read the README, and plan the move to a supported framework before an unpatched vulnerability makes the decision for you.
Comments
Post a Comment