On November 6, John Langford posted “The End of AI Frontiers” on his blog and walked out of Microsoft. Langford had founded AI Frontiers — a ~50-person research unit inside Microsoft Research — specifically in response to ChatGPT’s launch in 2023. The group spent three years building meaningful work: AutoGen, the multi-agent orchestration framework, plus novel architectures like the Full Bandwidth Transformer. All of it now belongs to a research org that no longer exists. And AutoGen, which developers are running in production today, officially receives no security fixes.
AutoGen Is Unmaintained — Check Your Stack Now
This is the part that matters most if you’re shipping code. The AutoGen README has carried an explicit disclaimer since October 16: “Microsoft is not providing bug fixes, security fixes, new features, or support for this project.” The last PyPI release, autogen-agentchat 0.7.5, dropped September 30. Nothing has shipped since. Known code execution vulnerabilities in AutoGen Studio remain unpatched, with a CVSS score of 9.8. Security researchers identified five critical issues in a single week earlier this year.
Here’s the quiet danger: the GitHub repository is not formally archived. Run curl https://api.github.com/repos/microsoft/autogen | python3 -c "import json,sys; d=json.load(sys.stdin); print('archived:', d['archived'])" and you get archived: false. That means automated dependency audits, Dependabot, and CI/CD pipelines won’t flag it. Developers who don’t manually check the README likely don’t know. Run this now:
pip freeze | grep -i autogenIf you see autogen-agentchat in that output, you need a migration plan.
What Killed It: Research Culture Meets Scaling Factory
The story of AI Frontiers’ death is straightforward, and Langford doesn’t sugarcoat it. The group was reorganized into Microsoft AI in early September 2026 — an org led by former Inflection AI executives whose priority, in Langford’s words, is “scaling, and aggressively not research.” Bigger models, longer training runs, more compute. AI Frontiers was built to do the opposite: find novel architectures that could make AI cheaper and more capable. The two cultures are genuinely incompatible, and research lost.
Langford called it “a bit of a train wreck.” He said he was “not comfortable being at a company where the research I do has no path to real use.” All other founding leaders reached the same conclusion and left. Microsoft has confirmed there will be no AI Frontiers internships next year, which is about as official a death certificate as you’ll get for a research lab.
Related: Thoughtworks Tech Radar Vol. 35: Agent Sandboxing Is Now Non-Negotiable
What Was Actually Lost
AutoGen gets most of the developer attention, but the Full Bandwidth Transformer is the quieter loss. The technique feeds each decoding step’s hidden state back as input — essentially giving models better memory of their own outputs. ByteDance experiments found it can match a model roughly twice its size. Langford made the economic case plainly: “How much is it worth to halve training and inference cost given that the industry has investments approaching a trillion dollars?” Microsoft, apparently, couldn’t answer that question in research’s favor.
The group also built Fara (a computer-use agent), the Next Latent training objective, and Dion3 (an improved optimizer). These are research artifacts now. No lab to continue them, no corporate mandate to productize them.
Where AutoGen Users Go From Here
You have three real options. Microsoft Agent Framework (MAF) is the official successor: a production SDK combining AutoGen and Semantic Kernel, available for Python and .NET since April 2026. It has Azure-native integrations and actual support. The catch: Anthropic support is still in beta, so if your stack isn’t Azure-centric, you’ll hit friction. AG2 is the community fork led by the original AutoGen paper authors, Chi Wang and Qingyun Wu. It’s backward-compatible with AutoGen v0.2’s GroupChat API, actively maintained (v1.1.2 landed October 3), and doesn’t require Azure. For teams already outside the Microsoft ecosystem, AG2 is the cleaner migration. LangGraph and CrewAI are solid third options if you’re willing to rewrite more of your orchestration logic.
Related: Microsoft APM: A Package Manager for AI Agents (2026)
Key Takeaways
- AutoGen (
autogen-agentchat) is officially unmaintained — no security fixes, no bug fixes. Migrate off it. - The GitHub repo is not archived, so dependency scanners won’t catch it. Audit manually with
pip freeze | grep -i autogen. - Microsoft Agent Framework is the official successor; AG2 is the community fork for provider-neutral teams. Both are actively maintained.
- AI Frontiers was dissolved because Microsoft AI prioritizes scaling over research — a culture clash that ended with every founding leader leaving.
- The Full Bandwidth Transformer and other research projects have no home. That work is likely done.
