Konstantin Semenenko
June 23, 2026
4
minutes read
Use Microsoft.Extensions.AI for the foundation, models, inference, embeddings, and data or index pipelines, and use Microsoft Agent Framework, built on top of it, for the next level up: agents, workflows, and multi-agent orchestration. Semantic Kernel was a mix of those two layers, which is exactly why Microsoft deprecated it.
Microsoft's .NET AI story finally has a clean shape: one foundation library, and one layer built on top of it. The confusing middle, Semantic Kernel, got folded into that layer, so the real question now is which of the two you reach for. Here's what Microsoft.Extensions.AI and Microsoft Agent Framework each do, and how to choose.
Use Microsoft.Extensions.AI for the foundation, models, inference, embeddings, and data or index pipelines, and use Microsoft Agent Framework, which is built on top of it, for the next level up: agents, workflows, and multi-agent orchestration. Semantic Kernel was a mix of those two layers, which is exactly why Microsoft deprecated it in favor of this split.
We build production .NET AI on both, so this is the mental model we actually use, not a spec sheet.
As two clear levels. At the root is Microsoft.Extensions.AI, the foundation every AI feature sits on. On top of it is Microsoft Agent Framework, for when the work becomes agentic. Semantic Kernel used to straddle both levels in a single SDK, and that overlap is the reason it's been superseded: Microsoft pulled the foundation down into Microsoft.Extensions.AI and pushed the agent parts up into Agent Framework, leaving two libraries with clear jobs instead of one that did a bit of everything.
The foundation: anything to do with talking to models and your data. That means inference and chat through a single IChatClient, embeddings through IEmbeddingGenerator, and the data and index pipelines behind retrieval and RAG, plus provider-agnostic middleware, telemetry, and retries. If what you need is new model support, an inference call, an embedding, or a data and indexing pipeline, this is the layer. It's the root the rest of the stack builds on.
The next level up: when something has to act, not just answer. That means agents that pursue a goal across multiple steps, use tools, and remember conversation state, plus workflows and multi-agent orchestration, with MCP, A2A, and checkpointing built in. It reached 1.0 GA in April 2026 as the successor to both Semantic Kernel and AutoGen, and it's built directly on Microsoft.Extensions.AI, so everything from the foundation carries straight through. If you need agents, workflows, or orchestration, this is the layer.
Because it was the mix of those two layers, and Microsoft decided the mix was the wrong unit. Semantic Kernel bundled foundation concerns (model access, embeddings, memory) together with agentic ones (plugins, planning, orchestration) in one SDK. The 2026 reorganization splits that cleanly: foundation goes to Microsoft.Extensions.AI, agentic goes to Microsoft Agent Framework, both built by the same teams. Semantic Kernel 1.x is still supported but in maintenance, fixes and security only, with all new work going into the two-layer model.
By what the work actually is. Need model support, inference, embeddings, or a data and index pipeline: Microsoft.Extensions.AI. Need the next level, agents, workflows, or multi-agent orchestration: Microsoft Agent Framework, which already includes the foundation underneath. Already on Semantic Kernel: move the foundational parts down to Microsoft.Extensions.AI and the agent parts up to Agent Framework, and don't start anything new on SK.
Start at the root and climb only when you have to. Build on Microsoft.Extensions.AI, and the day the problem turns into a goal-seeking agent or a multi-step workflow, add Microsoft Agent Framework on top. That's the whole decision, and it's a far simpler one than the framework debates of a year ago. The harder part is the judgment around it, what to let an agent do and what to verify, which is why we open-sourced our own .NET agent tooling; you can read about it in our open source. When you want a senior .NET team to build the AI, that's AI Dev Team.
