Microsoft's AutoGen left without security fixes as AI Frontiers lab dissolves

What happened

The Microsoft Research group behind the AutoGen agent framework has been dissolved, and the framework itself is no longer maintained: an October 16 change to the AutoGen repository states that Microsoft will not deliver bug fixes, security patches, new features, or support. According to a November 7 report on dev.to, the lab's demise was confirmed a day earlier by John Langford, who helped create the group, in a post announcing the end of AI Frontiers.

The end of AI Frontiers

Writing on hunch.net, Langford says AI Frontiers was a group of roughly 50 people set up inside Microsoft Research after ChatGPT arrived, to run agent research in a more coordinated way. Its work included the AutoGen orchestration framework and the Fara computer use agent, alongside fundamental research from its New York team: the Next Latent objective, the Dion3 optimizer (an improvement on Muon), and the Full Bandwidth Transformer.

In early September, according to Langford, the group was reorganized into Microsoft AI, whose leadership he saw as prioritizing scaling over research. He writes that he found new homes for his people and left the company, and that the group will not take interns next year. As dev.to notes, the post says nothing about the fate of AutoGen, Fara, or their maintainers. The Fara repository remains MIT-licensed and unarchived, with its last push on September 22.

AutoGen was already winding down

The lab's closure compounds a decline that started well before it. The final release of autogen-agentchat on PyPI is version 0.7.5, uploaded on September 30, 2025. The README now sends new users to Microsoft Agent Framework, its recommended replacement for production use.

The October 16 change came through pull request #8366, whose description says it makes explicit that the project is archived and no longer maintained by Microsoft; a later commit in the same pull request adds that AutoGen was never meant for production use, and the README warns that vulnerabilities may go unaddressed.

One detail will trip up automation: GitHub's API still reports the repository as unarchived, so dependency audits that check the archive flag will see nothing unusual. The README text is the clearest available signal.

Check which package you actually run

The AutoGen name covers more than one project. Microsoft's packages are hyphenated, such as autogen-agentchat and autogen-ext. The plain autogen package on PyPI belongs to the separate AG2 project, which says it receives only security and critical fixes, with active development in ag2. Meanwhile, pyautogen now acts as a proxy that installs autogen-agentchat. Teams should confirm which packages their environments hold before assuming they are affected by Microsoft's decision.

What migrating to Agent Framework involves

Microsoft Agent Framework 1.0 shipped on April 3, 2026 for .NET and Python, and the current Python release on PyPI is 1.25.0. Microsoft's migration guide says it was built by the core AutoGen and Semantic Kernel teams.

The port is more than a rename. According to the guide, AutoGen's AssistantAgent completes a single tool iteration by default, while Agent Framework's Agent keeps calling tools until it reaches a final answer, which can raise model usage per request. The Agent class is stateless between calls, so conversation history must live in an AgentSession. AutoGen teams map to typed Workflows with checkpointing and pause/resume behavior, an idea derived from AutoGen's experimental GraphFlow. And where AutoGen offered experimental distributed runtimes, Agent Framework currently focuses on single-process composition, with distributed execution still planned.

Provider coverage needs checking too: the guide's client table, dated October 23, still lists Anthropic and Ollama as planned, and agent-framework-anthropic on PyPI exists only as a beta pre-release (1.0.0b261106, uploaded November 7), which pip will not select by default.

Why it matters

Teams running AutoGen in production are doing so on software the vendor will not patch, including for security problems, and the usual archive-flag signals fail to reveal it. The successor is documented and actively released, but migration changes runtime behavior around tool loops, session state, and distributed execution rather than just class names. With the group that built the framework gone, the realistic paths are a deliberate port to Agent Framework, a move to the community-run AG2, or dropping the orchestration layer entirely — none of which is a drop-in upgrade.