If you sat down in mid-2026 to build a multi-agent system and someone told you to "just use AutoGen," you would have a worse afternoon than they intended. Type the name into GitHub and you get a 59,000-star repository. Type it into PyPI and you can install three subtly different things. Read the docs and one of them tells you, in its own README, that it is no longer being developed. The advice wasn't wrong so much as out of date: "AutoGen" stopped being a single project and became a three-way fork, and nobody updated the shorthand.

This is the part most comparison posts get backwards. They line up AG2 and AutoGen as if you were choosing between two competing tools the way you'd choose LangGraph or CrewAI or AutoGen. But AG2 and AutoGen aren't competitors that happen to look alike. They're the same project after a divorce, and the thing that actually splits them is hiding in plain sight: the version number.

One name, two architectures, three roadmaps

Here's the genealogy, because it's the whole story.

AutoGen began as a Microsoft Research project with a conversation-first design. Agents were ConversableAgent objects; you composed them into a GroupChat and they took turns talking. That was the v0.2 era, and it's the version nearly every tutorial, blog post, and Stack Overflow answer written before 2025 silently assumes.

In late 2024, the people who created AutoGen forked their own work out of Microsoft's repo and started AG2 — a community-governed continuation of that v0.2 design, with an open steering model and contributors from across the industry rather than a single corporate owner. AG2 kept the synchronous, conversational API and kept iterating on it; by mid-2026 it's on the v0.13 line and still shipping.

Then, in January 2025, the repo still named microsoft/autogen shipped v0.4 — not an upgrade but a ground-up rewrite. Gone was the synchronous chat loop; in its place an asynchronous, event-driven actor model, with a layered design separating a low-level core (message routing between agents-as-actors) from a higher-level AgentChat API. The motivation was real: the v0.2 design didn't scale cleanly to distributed, multi-process systems, and an actor model does. But it was a different program wearing the same name.

AG2 and AutoGen v0.4 are not two versions of one tool. They are two answers to "what should AutoGen have become," shipped by two different sets of people, under the same brand.

The twist: Microsoft's real bet is neither

If the story stopped there, you'd pick by taste — synchronous and community-run, or asynchronous and corporate. But there's a third path, and it's the one that explains why microsoft/autogen now reads as a maintenance-mode project.

In October 2025 Microsoft announced it was merging AutoGen and Semantic Kernel — its other agent framework — into a single Microsoft Agent Framework. That framework hit a release candidate in February 2026 and went 1.0 GA on April 3, 2026. It takes AutoGen's multi-agent abstractions and the actor/graph-workflow ideas from the v0.4 rewrite, and fuses them with Semantic Kernel's enterprise plumbing: session state, middleware, telemetry, multi-provider model support, and first-class .NET alongside Python.

Which means the v0.4 rewrite was less a destination than a stepping stone. Once the Agent Framework existed, both predecessors were quietly retired into maintenance — AutoGen's README now says as much in plain language. Microsoft didn't kill AutoGen; it absorbed the parts it wanted and let the original name go cold. This is the same consolidation move you can watch play out in the Semantic Kernel / AutoGen / Microsoft Agent Framework lineage — two frameworks becoming one.

So which do you install?

The decision tree is shorter than the history.

Starting fresh and want the conversational, v0.2-style API alive and improving? Install AG2 (pip install ag2). It's the only one of the three actively developing that design, the original authors are steering it, and the open governance means it isn't one company's strategy reversal away from being frozen. If your mental model of AutoGen is "agents that talk to each other in a group chat," AG2 is where that model still lives. It's also the natural base when you're sketching supervisor, swarm, or handoff orchestration and want to stay close to the classic primitives.

Want what Microsoft is actually building and supporting for production? Skip both legacy repos and go to the Microsoft Agent Framework. That's where the engineering investment, the enterprise features, and the .NET story are. The actor model you might have wanted from v0.4 is carried forward there, in a supported package.

Already running AutoGen v0.4, or studying the actor architecture? Stay, but know you're on a frozen branch. It works, it's interesting, and it's going nowhere new.

The trap is the middle option by default — typing pip install autogen because that's the name you remember, and landing on a maintenance-mode project or an ambiguously-owned package name. In 2026, the safest thing you can do with the word "AutoGen" is refuse to use it as an instruction. Name the version. The version is the fork.