This Month In Declarative Agents: MCP Gets Real

January was the exhale after Ignite. December dropped MCP support, a model upgrade, and enough news to keep everyone busy through the holidays. So January? Quiet on paper. But two things landed that are worth your time, and one of them fills a gap that was genuinely holding people back.

MCP Auth Finally Got a Real Tutorial

MCP support for declarative agents went GA in December, and the docs covered the basics. But if you wanted to see someone actually wire up Entra authentication, deal with the gotchas, and build something end-to-end? You were on your own.

A post on developerscantina.com changed that. It walks through building a declarative agent that consumes an Entra-authenticated MCP server: Agents Toolkit scaffolding, mcp.json configuration, OAuth setup through Azure API Management, and the specific quirks you’ll hit along the way.

What makes it worth reading is that it tackles the hard part. Connecting to an unauthenticated MCP server is straightforward. Connecting to one behind Entra, with proper token flows and API Management in front? That’s where most people get stuck, and this post doesn’t skip those steps.

It’s also the first in a series covering the same pattern with Copilot Studio and Microsoft Foundry. If you’re planning to expose internal APIs through MCP (and you should be), start here.

GPT-5.1 Rolled Out Under the Hood

This one flew under the radar, but it matters. The GPT-5.1 model upgrade (MC1194070) completed its rollout in mid-January. Every declarative agent now runs on GPT-5.1 with what Microsoft calls the “auto” architecture: the orchestrator picks between a fast chat model and an advanced reasoning model depending on query complexity. Simple lookups get the fast path. Multi-step reasoning gets the heavy model. You don’t configure this. It just happens.

In practice, agents feel snappier on routine questions and handle complex queries better. The reasoning over long SharePoint documents is noticeably more coherent, and response times on simple lookups dropped.

If you built a declarative agent before the holidays and haven’t tested it since, give it another run. The model underneath changed.

My Take on January

January was a transition month. The big announcements happened at Ignite, and the community needed time to absorb, experiment, and start writing. The Developers Cantina tutorial is exactly what I expected: someone taking the MCP announcement and building a practical, reproducible guide. More of that is coming.

The GPT-5.1 upgrade is the kind of platform improvement that doesn’t get blog posts but makes a real difference. Your agents got smarter without you changing a line of configuration. That’s the payoff of building on a managed platform.

My advice: go read that MCP tutorial and start thinking about which internal APIs could be exposed as MCP servers. The pattern is proven, the tooling works, and the auth story is finally documented.