Microsoft folded Semantic Kernel and AutoGen into one thing: Microsoft Agent Framework, which shipped a production-ready 1.0 for both .NET and Python and now carries native MCP and A2A support in the SDK. If you were choosing between the two old frameworks for tool integration, that decision is gone — and MCP is how you give an agent real tools. Here's the shape of it.
Native, not adapter
The distinction matters when you evaluate frameworks. Native support means the protocol ships in the SDK as a primitive; adapter support means a bridge exists but you maintain it. Agent Framework is native on both runtimes, so an MCP server is a first-class tool source rather than a community add-on. That's the same bar the Claude Agent SDK, LangGraph, and Pydantic AI clear — and it's the property to look for when you pick a stack.
Attaching a server
The pattern is the same one every native client follows: point the framework at a server, let it discover the tools, and hand them to an agent. A server can be a local stdio process or a remote HTTP endpoint. In Python the flow is to construct an MCP client for the server, list its tools, and pass them into the agent's tool set; in .NET the DI-friendly ModelContextProtocol package plugs the same tools into an AIAgent. Either way the agent treats an MCP tool exactly like a function you wrote by hand — the framework converts each tool's JSON schema into its own tool type.
Progressive tool disclosure
The interesting 2026 addition is agent-framework-core's progressive MCP disclosure: an agent can discover, load, and unload MCP tool schemas on demand while keeping the allowed_tools boundary intact. This is the fix for the problem every large deployment hits — dozens of servers, hundreds of tools, and a context window that fills with schemas the agent never calls. Instead of front-loading every definition, the agent pulls schemas in as a task needs them and drops them after. It's a direct lever on cost and latency; we go deeper in MCP progressive tool disclosure and agent token budget control.
Where it pays off
Multi-agent systems are the natural home. A supervisor routes work to specialists that each own a different server — one on GitHub, one on a Postgres database, one on Slack — without sharing code, and progressive disclosure keeps each agent's context lean. That separation is what keeps a large system auditable and lets you swap a tool without touching agent logic, the theme of our multi-agent orchestration patterns.
Where to go next
New to the protocol? Start with what an MCP server is, then decide stdio versus HTTP in transports compared. Building the servers yourself follows the Python or C# tutorial, and everything you connect should pass MCP security best practices. Browse the ML engineer loadout for a vetted first stack.