This Month In Declarative Agents: Agents Stopped Just Talking

For two years, talking to a declarative agent meant reading. You asked, it answered in text, and if you were lucky you got a citation. June changed the deal. Agents can now hand you something to look at and click, not just a paragraph to parse. The whole month leaned in one direction: interactive UI, and a single integration layer underneath it.

Copilot chat grew widgets

The headline came from Microsoft with SharePoint Copilot Apps, interactive UX components that render right in the Copilot canvas. They run on the MCP Apps model, they sit on top of declarative agents, and you build them with SPFx using whatever JavaScript framework your team already knows. Approve an expense, check remaining leave, book a desk, all without leaving the chat. That’s a real shift from “agent that answers” to “agent that does.”

Here’s the part I loved: the community had the full build guide ready the same week. Andrew Connell published a four-part walkthrough on voitanos.io that goes from concept to keypress. It opens with what MCP Apps actually are and when to reach for them over adaptive cards, then adds an MCP server with a React widget to a declarative agent project, wires that server into the agent with a RemoteMCPServer plugin manifest and rewritten instructions, and finishes by folding the whole thing into the Agents Toolkit F5 debug loop so one keypress launches everything.

When a platform announcement and a complete, testable walkthrough land within days of each other, that tells you the tooling is real, not a slide.

One MCP server, every surface

The other story was quieter, and probably more useful for how you plan your work. Matteo Pagani ran a two-post arc on developerscantina.com showing that the authenticated MCP server you build once can power a lot more than one agent. First he turned an Entra-protected .NET MCP server into a federated Copilot connector that serves live data into Researcher and Copilot Chat with nothing ingested into the tenant. Then he took the exact same server and OAuth registration and reused them as a Cowork plugin, with no new app registration to create.

Read those back to back and the lesson is hard to miss. Declarative agents, federated connectors, and Cowork all speak MCP, and they can share one auth setup. Build the server and the OAuth connection once, point each surface at it, done. The official Copilot Developer Camp added Cowork labs this month too, so the hands-on learning path is growing in the same direction.

My Take on June 2026

June was the month agents got a face and a shared backbone. Interactive widgets through MCP Apps closed the distance between “looks like a chatbot” and “feels like a tool,” and the MCP-everywhere pattern means the work you do for one surface stops being throwaway.

If you build one thing this summer, make it an authenticated MCP server with a couple of real widgets. That single artifact now buys you an interactive declarative agent, a federated connector, and a Cowork plugin. A month ago that was three projects. Now it’s one.

The direction is obvious, and I’m here for it. The platform is treating MCP as the connective tissue across every Copilot surface, and it’s finally letting agents show instead of only tell. Plan your next agent like UI and reuse are the defaults, because as of this month, they are.