The lab behind AutoGen is gone, and John Langford says scaling killed it
Microsoft's agent research group just got a eulogy from one of its founders. John Langford, the machine learning theorist who runs the hunch.net blog, posted "The End of AI Frontiers" on Nov 6, and it was the top AI story on Hacker News Saturday with 251 points.
AI Frontiers was a roughly 50-person unit inside Microsoft Research, set up in 2023 to run research as organized efforts instead of lone projects. Langford writes that much of its work was "figuring out how to make agents work", naming the AutoGen orchestration framework and the Fara computer-use agent. In early September it was reorganized into Microsoft AI, the larger group led by ex-Inflection leadership. Langford's verdict on that culture is blunt: "scaling, and aggressively not research", and "plenty of evidence that they just did not really want us."
Langford left, says every other founding AIF leadership member reached a similar conclusion, and says there will be no AIF internships next year. Going back to MSR wasn't attractive either: the post argues research there has no path into products, and that the basic deal of "do good work and good things happen" has broken down. This is one person's account, and we found no Microsoft statement.
Why agent builders should care: AutoGen is one of the most used multi-agent frameworks, and its research home just dissolved into an org that, by this account, doesn't want research. Microsoft has already been steering developers toward its newer Agent Framework, so the practical question is who keeps AutoGen alive and how fast. If you depend on it, watch the commit graph, not the press.
There's also a useful argument inside the complaint. Langford points to ByteDance experiments suggesting a variant of the Full Bandwidth Transformer can match a model twice its size, and asks what halving training and inference cost is worth in an industry spending close to a trillion dollars. Scaling is predictable, research isn't, and big labs keep picking predictable. That's how the next efficiency trick ends up at a smaller lab.
Post: https://hunch.net/?p=13764006
← Back to all articles
AI Frontiers was a roughly 50-person unit inside Microsoft Research, set up in 2023 to run research as organized efforts instead of lone projects. Langford writes that much of its work was "figuring out how to make agents work", naming the AutoGen orchestration framework and the Fara computer-use agent. In early September it was reorganized into Microsoft AI, the larger group led by ex-Inflection leadership. Langford's verdict on that culture is blunt: "scaling, and aggressively not research", and "plenty of evidence that they just did not really want us."
Langford left, says every other founding AIF leadership member reached a similar conclusion, and says there will be no AIF internships next year. Going back to MSR wasn't attractive either: the post argues research there has no path into products, and that the basic deal of "do good work and good things happen" has broken down. This is one person's account, and we found no Microsoft statement.
Why agent builders should care: AutoGen is one of the most used multi-agent frameworks, and its research home just dissolved into an org that, by this account, doesn't want research. Microsoft has already been steering developers toward its newer Agent Framework, so the practical question is who keeps AutoGen alive and how fast. If you depend on it, watch the commit graph, not the press.
There's also a useful argument inside the complaint. Langford points to ByteDance experiments suggesting a variant of the Full Bandwidth Transformer can match a model twice its size, and asks what halving training and inference cost is worth in an industry spending close to a trillion dollars. Scaling is predictable, research isn't, and big labs keep picking predictable. That's how the next efficiency trick ends up at a smaller lab.
Post: https://hunch.net/?p=13764006
Comments