For two years, a developer building agents on Microsoft's stack faced a fork with no good signpost. Semantic Kernel was the enterprise SDK — typed, .NET-first, the thing quietly powering Microsoft 365 Copilot. AutoGen was the Microsoft Research darling — a multi-agent conversation framework with the bigger community and the more interesting demos. They overlapped, they competed, and Microsoft maintained both. Choosing meant guessing which one Redmond would still care about in a year.
Microsoft just answered the question by retiring both.
The merger, stated plainly
The Microsoft Agent Framework (MAF) is the single successor to Semantic Kernel and AutoGen, built by the two teams combined. It went to public preview on October 1, 2025, hit a Release Candidate on February 19, 2026, and reached 1.0 general availability on April 3, 2026, with a long-term-support commitment for .NET and Python.
The design is the merger made literal: AutoGen's lightweight single- and multi-agent abstractions, wrapped in Semantic Kernel's enterprise plumbing — session-based state, type safety, filters and middleware, and built-in OpenTelemetry. On top sits a graph-based workflow layer for explicit multi-agent orchestration (sequential, concurrent, handoff, group chat, and the Magentic pattern). It speaks MCP for tools and A2A for talking to agents in other runtimes.
The most-cited fact about Microsoft Agent Framework is that it's new. The most useful fact is that it's a tombstone: a 28k-star SDK and a 59k-star research project both now point at an 11k-star successor.
What this does to your decision
The practical consequence is that the old comparison is dead. You no longer weigh Semantic Kernel against AutoGen on the merits, because Microsoft has placed both into maintenance mode — bug fixes and security patches, with new feature work flowing to MAF.
So the rule is blunt. New project on Microsoft's stack? Build on Agent Framework. Existing app on SK or AutoGen? You don't have to move this quarter, but you should plan to — Microsoft ships dedicated migration guides precisely because the move is a port, not a flag flip. Semantic Kernel's safety net is roughly a year of critical fixes past GA. "We'll migrate later" is a real decision with a real deadline attached.
A note on "RC/GA stable APIs," since it's easy to over-read: that promise covers the surface — the v1.0 API won't churn underneath you. It says nothing about your migration cost from SK or AutoGen, which are different enough frameworks that there's a guide for each.
Where it wins, and where it doesn't
Against the Python-native field — LangGraph, CrewAI, and newer entrants like Agno — MAF has two genuine differentiators, not marketing ones.
The first is language parity. MAF treats .NET (C#) and Python as first-class, GA'd together. LangGraph and CrewAI are Python-only. If your production system is a C# shop, MAF is the only mainstream agent framework that doesn't make you the second-class citizen. (The honest caveat: samples and ecosystem density still skew Python early, so verify the specific feature you need has parity before betting a C# service on it.)
The second is identity. Microsoft Entra Agent ID gives each agent a real directory identity — a service principal you can grant scopes to, audit, and revoke — which is the kind of governance control enterprises actually block deployments over, and which the Python frameworks leave to you to bolt on.
The catch is the same thing that makes those features possible: Azure gravity. The MAF core is open-source and model-agnostic; you can point it at OpenAI, Anthropic, or a local model. But the value that justifies choosing it — Azure AI Foundry deployment, Entra identity, Azure Monitor traces, Cosmos-backed state — compounds inside Azure and thins out elsewhere. If you're already on Azure, the marginal cost of MAF is close to zero and the integration is the whole point. If you're picking it up outside Azure, you're carrying a framework whose best parts are dimmed, and a self-hosted LangGraph on a Postgres checkpointer may cost you less in the end.
That's the real shape of the choice now. Within Microsoft's world, the framework question is settled by fiat — MAF, and only MAF, going forward. The open question isn't which Microsoft framework; it's whether you want to be in Microsoft's world at all. The protocol layer (MCP and A2A) is Microsoft's own answer to that — interop good enough that the lock-in lives in the conveniences, not the wire format. Which is exactly where a vendor wants it.
