What Microsoft Actually Combined in One Stack
I've often seen questions like: is this Microsoft's fourth agent framework or the fifth? In reality, no. We're not talking about another new zoo of tools, but a proper consolidation of two lines that previously existed side-by-side: AutoGen for orchestration and Semantic Kernel for enterprise-level scaffolding.
On April 4, I dug in to see what was marketing and what was a real shift, and the picture is quite down-to-earth. Microsoft Agent Framework 1.0 was released as a production-ready SDK for .NET and Python, with a clear long-term bet on it. It was announced back in fall 2025, with RCs in February 2026, so it's no longer just a demo story.
What caught my eye wasn't the word 'Framework' but the fact that they finally stopped making people glue half the system together manually. It offers a single runtime, a unified model for multi-agent orchestration, graphs, streaming, checkpointing, and human-in-the-loop. It's not a revolution. But it finally looks like a tool you can bring to production, not just to presentations.
The Catch with Azure: Is There Vendor Lock-in?
I share the skepticism about vendor lock-in. Microsoft has, let's say, an earned reputation on this matter. But in this case, I wouldn't reduce it all to "please, use only Azure AI Foundry."
The Agent Framework is open-source under the MIT license and supports not only Azure but any OpenAI-compatible endpoint. Plus, you can run local models via Ollama and host everything in ASP.NET Core without being tied to the Microsoft cloud. For those designing AI solution architectures and thinking ahead about how to avoid surgically removing platform dependencies later, this is a good sign.
The shift in August 2025 with the Azure OpenAI v1 APIs is also particularly important. Back then, Microsoft clearly started removing unnecessary Azure-specific overhead and improving compatibility with OpenAI clients. I read this as them realizing that the winning stack for AI adoption isn't the most closed one, but the one that gets in the developer's way the least.
What About RAG? No Longer a Half-Day Quest
This is where Microsoft really struggled with DX for a long time. Anyone who tried to set up RAG via Azure AI Search in 2023-2024 remembers the circus with indexes, pipelines, manual chunking, and other joys. Not impossible, but you couldn't call it user-friendly.
The picture is better now. The Foundry Agent Service and the new agent stack offer a much clearer File Search approach: upload files, set up a vector store, connect it to an agent, and the service handles chunking, embeddings, keyword + semantic search, and reranking. For enterprise RAG, this at least lowers the barrier to entry.
That said, I wouldn't pretend the magic bullet has been found. If you have a complex domain, non-standard access rights, heavy documents, multilingual search, or explainability requirements, the "just throw files at it" abstraction quickly falls apart. But as a foundational layer for AI integration, it's something you can present to your team without embarrassment.
Who Really Benefits from This?
The winners are teams that need a working agent system for business processes, not a research sandbox. This is especially true where you need to quickly build an internal assistant, a RAG system for documents, an operator copilot, or AI automation on top of a CRM, helpdesk, and knowledge base.
The losers, ironically, aren't Microsoft's competitors, but old homegrown setups. The ones where orchestration lived in one place, retrieval in another, and the glue code took up more space than the actual logic. I've seen such contraptions more than once, and maintaining them is a pain.
At Nahornyi AI Lab, this is usually where we step in: when the goal is not just to try out an agent, but to build AI automation with a solid architecture, without unnecessary cloud dependencies or production surprises. Sometimes the Microsoft stack is a perfect fit. Sometimes it's not, and then I honestly look at other options.
This breakdown was written by me, Vadym Nahornyi, from Nahornyi AI Lab. I don't just repeat press releases; I build and test things like this with my own hands in real use cases: RAG, agentic pipelines, AI architecture, and integrating artificial intelligence into business processes.
If you want to discuss your case, order custom AI automation, create a bespoke AI agent, or build an n8n automation with an LLM on top of your data, contact me. We'll figure out if the Microsoft stack is needed here, or if there's a shorter, cheaper path.