AutoGen vs LangGraph in 2026: Microsoft Has Already Answered This

Most comparisons of these frameworks argue about orchestration models. That argument is worth having, and it is the second question. The first one is which of them is still being built, and on that the answer is unusually clear.

Repository activity, read from the GitHub API on 9 September 2026:

AutoGen has the most stars of any of them and is the only one that has stopped moving. Five months without a push, with 1,063 open issues. It is not archived, and nobody has announced anything, which is exactly why a star count is such a poor proxy for whether a project is alive.

Microsoft has already said what replaces it

This is not a reading of tea leaves. Microsoft's Agent Framework documentation says it directly:

Semantic Kernel and AutoGen pioneered the concepts of AI agents and multi-agent orchestration. The Agent Framework is the direct successor, created by the same teams.

And, in the same section:

In short, Agent Framework is the next generation of both Semantic Kernel and AutoGen.

There is a published migration guide from AutoGen. Agent Framework combines, in Microsoft's words, "AutoGen's simple abstractions for single- and multi-agent patterns with Semantic Kernel's enterprise-grade features such as session-based state management, type safety, filters, telemetry", and adds graph-based workflows on top.

So the question "AutoGen or LangGraph for a new project" has a component most comparisons miss: one of the two options is a framework whose own authors have shipped its replacement and told you to move.

That does not make AutoGen useless. Existing systems keep working, the ideas were good enough to be carried forward wholesale, and a framework with no commits is not a framework with no users. It does mean that choosing it fresh in 2026 means choosing a codebase that receives no fixes.

The fork worth knowing about

If you want the AutoGen programming model and would rather have a community project than a Microsoft one, AG2 is the live option. Its repository describes it as "AG2 (formerly AutoGen)" and it was pushed on 8 September 2026.

The size difference matters and cuts both ways: 4,914 stars against the original's 60,889. Almost none of AutoGen's accumulated attention followed the fork. You get active maintenance and a much smaller community, fewer answered questions on Stack Overflow, and fewer blog posts describing your exact problem.

The actual design difference, once you get past maintenance

Assume for a moment all three were equally maintained. They would still suit different work, and the difference is not features. It is what the framework thinks a program is.

AutoGen thinks a program is a conversation. Agents talk, and control flow emerges from who speaks next. You configure participants and let the routing happen. This is genuinely fast to prototype, and it is why the model was worth carrying into Agent Framework. It also means that when the system does something strange, the explanation is distributed across a transcript rather than located in a line of code.

LangGraph thinks a program is a graph with typed state. You draw the nodes and edges, and state moves through them with a shape you declared. Control flow is not emergent, it is the thing you wrote. That costs more up front, and it repays when you have to debug a failure at three in the morning or explain the system to somebody who was not there when it was built.

CrewAI thinks a program is a team. Agents are roles with goals, and tasks get assigned. It is the fastest of the three from nothing to something running and the least explicit about what happens in the middle.

The trade is the same one that runs through all of software: emergent behaviour is quick to build and hard to reason about, explicit structure is slow to build and easy to reason about. Nothing about agents changes it.

Choosing, in 2026

Start on LangGraph if the work is long-running, has to survive a restart, needs human approval in the middle, or will be maintained by somebody other than its author. The typed state and explicit edges are the reason.

Start on Microsoft Agent Framework if you are in the Microsoft ecosystem, want the AutoGen abstractions with enterprise features attached, or are already on Semantic Kernel. It is the successor to both by Microsoft's own account, ships in .NET, Python and Go, and is actively developed.

Start on CrewAI if you want role-based collaboration and speed to first result, and you can accept that control flow is less visible.

Start on AG2 only if you specifically want AutoGen's model and prefer a community fork to a Microsoft one, with clear eyes about the community size.

Do not start on microsoft/autogen. Not because it is bad, but because its authors have published its successor and a guide for leaving it.

If you already run AutoGen in production, none of the above is a reason to rewrite this quarter. Read the migration guide, note that no fixes are coming, and plan the move on your own timetable.

What none of these decide

A framework choice matters less than most comparison articles imply, including this one. All three orchestrate model calls and tool use. None of them makes the model better at the task, and the failure modes that actually hurt in production, prompt injection through untrusted inputs, tool calls that do more than expected, and cost that climbs quietly, are properties of running agents at all rather than of which library sequences them.

If you are choosing between these because you are building a multi-agent development workflow, the multi-agent software development workflow piece covers the shape of that work, and the best AI coding agents roundup covers the tools that already implement it, which is often the better answer than building your own orchestration.

FAQ

Is AutoGen still maintained?

microsoft/autogen was last pushed on 15 April 2026 and carries over a thousand open issues. It is not archived, but it is the only framework in this comparison that has not shipped in months. Microsoft's own documentation calls Agent Framework "the direct successor" and "the next generation of both Semantic Kernel and AutoGen", and publishes a migration guide from AutoGen.

Should I start a new project on AutoGen?

Not on microsoft/autogen. Microsoft points new work at Microsoft Agent Framework, which is built by the same teams and combines AutoGen's agent abstractions with Semantic Kernel's enterprise features. If you want the AutoGen programming model specifically and prefer a community project, AG2 is the active fork.

What is AG2?

The community fork of AutoGen, described on its own repository as "AG2 (formerly AutoGen)". It has far fewer stars than the original, about 4,900 against 60,900, and unlike the original it is being actively developed.

What is the real difference between AutoGen and LangGraph?

The mental model. AutoGen organises work as a conversation between agents, so control flow emerges from who speaks next. LangGraph organises work as a graph with typed state, so control flow is something you draw. Emergent routing is faster to prototype; explicit routing is easier to debug and to reason about a year later.

Where does CrewAI fit?

The role-based option: you define agents as roles with goals and let them collaborate on tasks. It is the fastest of the three to get something running and the least explicit about control flow. On repository activity it sits with LangGraph rather than with AutoGen.

Which has the most GitHub stars?

AutoGen, at about 60,900, ahead of CrewAI at 58,300 and LangGraph at 41,300. That ranking is the reason this comparison is worth doing carefully: stars record accumulated attention, not current maintenance, and here the two disagree completely.

When the whole comparison is the wrong question

Sometimes the answer to "AutoGen or LangGraph" is neither.

If your task is a sequence of steps you can write down, write the function. Microsoft's own documentation puts it more bluntly than any of these frameworks' marketing: "If you can write a function to handle the task, do that instead of using an AI agent." Multi-agent orchestration earns its complexity when steps are genuinely open-ended, when several specialisations really do need to coordinate, or when a human has to approve something in the middle. A pipeline with four known stages is a script, and a script does not need a framework, a graph or a crew.

The corollary is where this article could be wrong: if you are prototyping, and the point is to find out what the system should do rather than to run it reliably, then the emergent, conversational model is the better tool and the explicit graph is premature engineering. Prototype in whatever gets you an answer fastest, and choose the framework when you know what you are building.

Sources

Repository figures read from the GitHub API on 9 September 2026. Microsoft's position quoted from the Microsoft Agent Framework overview. Project repositories: microsoft/autogen, langchain-ai/langgraph, AG2, CrewAI.