AutoGen

AutoGen taught many people what a multi-agent conversation looks like, and Studio is still a nice way to show a manager what agents do. But the vendor has stopped adding features, and a framework is a dependency you live inside: every model API change and every Python release becomes your problem when nobody upstream is picking them up. Nobody should start a project on a framework whose own readme says that.

If you are already on it, the pain is proportional to how much of agentchat you used. Pick LangGraph if you need a runtime with checkpoints, smolagents if you want something small enough to read.

The risk is written in the repository. AutoGen is now in maintenance mode, will not receive new features or enhancements, and is community managed going forward. In practice that sentence means security fixes depend on volunteers, a dependency advisory waits on whoever notices, and the issue tracker becomes a place where questions age. A framework that executes generated code is not a place to accept that.

Budget the migration now, while the people who wrote the integration still work here. What it does right: the 0.2 to 0.4 migration guide is published, so a team leaving knows exactly what its code is built on and what the successor expects.

The 0.4 architecture is layered. 1. autogen-core: message passing, event-driven agents, a runtime that is local or distributed. 2. autogen-agentchat: the high-level API most users touched. 3. autogen-ext: model clients and capabilities. 4. Studio: a GUI over the rest. Generated code executes in Docker containers, which is the principled choice and one most agent frameworks still leave to the user. No benchmark is claimed, and none is needed for a framework whose claims are structural.

The consequence is that the design is worth reading even by teams that will not adopt it. The observation: a redesign this clean usually precedes a successor, and here it did.

A Microsoft sunset: no shutdown, just a readme pointing new users to Microsoft Agent Framework, the successor with stable APIs and long-term support promised. This is what an open-source sunset looks like when the sponsor can afford to keep the lights on for stragglers: the repository stays up, the momentum leaves. Sixty thousand stars are an audience being migrated, not an asset being sold, and the migration path is the product.

No acquirer for a project Microsoft already owns, and no pivot, because the pivot has a name. Position: closed; follow the users to the successor and price the migration into any estimate.

The demo is AutoGen Studio, where non-engineers wire agents together in a browser. As a dependency: Python and C# coverage matches our teams, the price is $0, onboarding is a pip install, and there is nobody at the vendor accountable for a bug or a CVE. A framework nobody owns is a liability on our side of the ledger, and it is sixty engineers' liability, not one person's.

Any team using it needs a migration ticket with a quarter attached, and any new project needs a reason in writing. Not yet, and not again; fund the move.

Everything is readable under MIT, MCP arrives through McpWorkbench, and autogen-ext model clients cover local endpoints, so an Ollama box is a config change. The interesting part is that the fork test has already been run: in November 2024 the original maintainers took the 0.2 line to ag2ai/ag2, Apache-2.0, pip install ag2, and kept shipping. A fork surviving the vendor is exactly my longevity test, passed early and in public.

So the choice is which lineage to read: Microsoft's 0.4 runtime, frozen and clean, or AG2, moving and familiar. Either way the code is mine, which is more than most of this board offers.