This Month In Declarative Agents: MCP Apps Goes Live

April was a quieter month than March on the community blog side, but the official side made up for it. MCP Apps left preview, the launch partners showed up, the new lab landed, and the docs team quietly killed one of the most common anti-patterns in declarative agent design. Two themes, one clear direction.

MCP Apps Stopped Being a Preview

On April 7, Microsoft published MCP Apps now available in Copilot chat. This is the GA story for the interactive UI work that started in March. Same idea (HTML widgets rendered inline in Copilot chat through MCP Apps or the OpenAI Apps SDK), but now with launch partners shipping in the Microsoft 365 Agent Store: Power Apps in public preview, plus Adobe Express, Coursera, Figma, and monday.com all live. That partner list matters more than it looks. It is the first time you can point at agents in Copilot chat and say, “those are not Microsoft demos, those are real third-party experiences.” The pattern is real, the rendering surface is stable, and the governance story (sandboxed iFrame, Entra SSO, admin policies) is the same as any other declarative agent. Build with confidence.

The post is also where Microsoft formally introduced the GitHub Copilot CLI skill for the Microsoft 365 Agents Toolkit. Install the work-iq plugin, then describe the agent you want in natural language, and the skill scaffolds the MCP server, wires up authentication, generates the widgets, and deploys. If you have not played with this yet, it is the fastest path I have seen from “I have an idea” to “there is a working declarative agent in my tenant.” Pair it with the official MCP Apps samples (Field Service Dispatch, Trey Research HR Consultant, Employee Training) and you have a credible starting point for almost any agent.

The Copilot Dev Camp also extended its end-to-end MCP track with a new Lab E11 on MCP Apps, merged on April 20. It picks up where E10 (MCP auth) leaves off and walks through wiring an MCP App into a declarative agent with the Agents Toolkit. If you went through E08 to E10 in March, finishing E11 in April closes the loop. You now have a complete, opinionated path from “MCP server” to “interactive agent in chat” that you can hand to any developer on your team.

The Docs Killed an Anti-Pattern

Buried in a single PR on April 27, the docs team updated the declarative agent instructions guidance to call out a pattern that has been quietly hurting agents in production: cramming everything into the instructions field. The new note is short, but the message is clear. Move tool selection logic, response formatting rules, and domain knowledge out of the instructions and into the actual primitives that exist for them (tool descriptions, plugin manifests, embedded knowledge, capabilities). Use instructions for persona, scope, and tone. That is it.

This is the kind of guidance that usually lives in a community post six months after enough people get burned. Seeing it land in the official docs means the platform is mature enough to have opinions about how to build on it. If your agent’s instructions are over a few hundred tokens and you are debugging why it picks the wrong tool, this PR is the one to read.

My Take on April 2026

April was a finishing month. March set the architecture (interactive UIs, MCP Apps, GPT-5.2). April locked it in with GA, partner launches, a finished lab track, and a docs correction that tells you how to actually use the thing well. There was no new shiny capability to chase, just the previous month’s bets becoming real.

The bigger signal sits next to the agent story. The April 29 Teams CLI and teams-dev agent skill post is not strictly about declarative agents, but it is the same playbook: an AI coding agent that takes care of the registration, manifest, and infra work so you can stay focused on logic. Pair that with the work-iq CLI skill on the declarative agents side and the pattern is obvious. The default way you scaffold M365 agents, declarative or otherwise, is going to be “describe it to a coding agent and let it set everything up.” If you have not adapted your workflow for that yet, May is a good month to start.