AutoGen is in maintenance mode. Where to migrate now.

AutoGen is not getting new features. The project README states it plainly: "AutoGen is now in maintenance mode. It will not receive new features or enhancements and is community managed going forward."

That is a clear signal. Most write-ups stop at repeating it. The useful question is what the move costs. Microsoft's own docs answer that, and the answer is not "swap the import".

What is Microsoft Agent Framework?

Microsoft Agent Framework is an SDK for building agents and workflows. It covers .NET, Python and Go. Microsoft's Agent Framework overview calls it the direct successor to both AutoGen and Semantic Kernel, built by the same teams.

That is the part to grasp. It is a merge, not a rename. It takes AutoGen's simple agent parts. It adds Semantic Kernel's enterprise plumbing, such as session state, type safety and telemetry. Then it puts graph-based workflows on top.

AutoGen is Microsoft's older multi-agent framework. It is built around an event-driven core and a high-level Team object. Semantic Kernel is Microsoft's older SDK for wiring models into apps. Agent Framework is the next step for both.

What actually changes when you port

The AutoGen migration guide lists four changes that matter. None of them is skin deep.

Area	AutoGen	Agent Framework
Orchestration	Event-driven core plus a high-level Team	Typed, graph-based Workflow routing data along edges
Tools	Functions wrapped with FunctionTool	@tool decorator, schemas inferred automatically
Agent turns	AssistantAgent is single-turn unless you raise max_tool_iterations	Agent is multi-turn by default until it can answer
Runtime	Embedded plus experimental distributed	Single-process composition today; distributed planned

The surface looks alike. Both build an agent from a model client, pass instructions, and attach tools. A hello-world agent looks much the same in both.

The first row is where the work is. An event-driven team is a chat. A typed workflow is a graph. It has edges you declare, and steps that fire when their inputs are ready. You cannot map one to the other line by line. You redraw the flow.

The state change that will bite you quietly

Here is the change that throws no error. In AutoGen, AssistantAgent keeps chat history as part of its state. In Agent Framework, Agent is stateless. It keeps no history between calls.

Your code still runs. Your agent just forgets the last turn. To keep a chat going, you attach AgentSession yourself.

MCP went stateless. Your server still has state. The lesson repeats here. Frameworks keep pushing state out of the default path and back onto you. Hidden state is what makes agent systems hard to reason about. Plan for it, rather than meeting it in a support ticket.

The other quiet change is loop behavior. Agent Framework agents keep calling tools until the job is done. They do not stop at a count you set. Microsoft says built-in safety guards prevent endless loops. That is kinder for simple tasks and harder to price. Watch your token bills for the first week.

Check your model providers before you plan anything

This one is a real trap. It comes from Microsoft's own docs disagreeing with themselves.

The migration guide's provider table marks Anthropic and Ollama clients as planned. It marks caching as planned too. The overview page, updated later, lists Anthropic and Ollama among supported providers.

Both pages are official. They cannot both be current. So check the provider list in the agent-framework repository. Then run a spike against the provider you actually use. Do that before you commit a sprint. Consider what happens if you plan a two-week port around a table that was right in April.

Do you run Ollama on your own hardware? Do you route through Anthropic for cost or region reasons? Then that spike is the first job this week. The cheapest AI API is not the cheapest to run. A framework that cannot reach your cheapest provider quietly deletes the saving.

Language support is not even

Agent Framework ships for .NET, Python and Go. The three are not equally far along. The Go version is in public preview. Declarative agents, RAG, CodeAct and functional workflows are not there yet.

If your service is Go, read that list against what you use. Do it before you plan the move. Python and .NET are the fuller paths today.

When should you leave Microsoft's stack entirely?

A port is a rewrite. So it is the cheapest moment you will ever get to switch vendor. Three cases make that worth a look.

1. You never used the Microsoft parts. Say your AutoGen app is one agent calling three tools against OpenAI. Then you have no Microsoft tie to keep. Any framework will take you.
2. You need work spread across machines now. Agent Framework runs in one process today. Spreading it out is planned, not shipped.
3. You want a different control model. A graph framework such as LangGraph gives you nodes and edges. A role framework such as CrewAI gives agents job titles. Pick the one that matches how you describe the problem out loud.

Take care with vendor comparisons here. That includes LangChain's own comparison page. It sums up what the maintenance notice changed. It is also written by a rival. Read it for the facts, not the verdict.

Data boundaries, for European teams especially

Agent Framework's docs carry a plain warning about third-party systems. Non-Azure models, servers and agents count as non-Microsoft products. They fall under their own license terms. Microsoft says you carry the usage and the costs.

The line that matters for a European team is about data flow. Microsoft says it is your job to manage whether your data leaves your Azure compliance and geographic boundaries. Do you route an EU workload through a third-party model endpoint? Then that is a review item, not a footnote.

The decision, in one line

Are you on AutoGen and staying with Microsoft? Start the port now. Budget a redesign for any multi-agent team, not a port. If your app never used Microsoft-only pieces, treat this as a free window to weigh LangGraph or CrewAI instead.