This Month In Declarative Agents: Interactive UIs Land

March was the month agents stopped just talking and started showing. Three things shipped that, together, change the default architecture for any new declarative agent project: interactive UIs in chat, a major toolkit release, and a model upgrade that makes everything work better.

MCP Apps Changed What Agents Can Do

On March 9, Microsoft announced that declarative agents can now render rich, interactive UI directly inside Copilot chat through MCP Apps. Not adaptive cards. Full HTML in a sandboxed iFrame: forms, dashboards, maps, approval panels, sortable tables.

Before this, your agent found an answer and responded with text. The user then left chat to actually do something with it. Now the action happens inline. An expense approval agent surfaces a table of pending items with approve and reject buttons. A field service agent shows a map with dispatched routes. The distance between “agent found the data” and “user took action” collapsed.

Copilot Dev Camp shipped a new lab before the month ended that walks through the full pattern: building an MCP App with request forms, approval panels, and status timelines, then wiring it into a declarative agent with the Agents Toolkit. If you want to build your first MCP App, start there.

The Toolkit Caught Up

Agents Toolkit v6.6.0 shipped March 9 and it’s the biggest release since the rename from Teams Toolkit. Three changes matter for declarative agent builders.

Embedded knowledge arrived. You can now package curated content directly with your agent instead of pointing at SharePoint or OneDrive. Product specs, compliance docs, FAQs, all shipped as part of the agent package and grounded first by the model. This is the kind of feature that makes agents feel like real products instead of chat wrappers.

MCP integration hit GA. The toolkit’s MCP support for declarative agents moved from preview to production-ready. Combined with MCP Apps, the path from “I have an MCP server” to “it’s a deployed declarative agent” is now fully supported and stable.

61 templates got cut. Sounds alarming, but it’s the right call. The toolkit was carrying legacy templates for patterns nobody should start fresh with. What’s left is focused: declarative agents, API plugins, MCP servers. Less noise, clearer direction on what to build.

A detailed review on voitanos.io digs into the release, including the hotfix history since v6.4.0 and what changed under the hood. Worth reading if you’re maintaining toolkit projects.

The Model Got Smarter

Microsoft upgraded all declarative agents to GPT-5.2 through March, completing the rollout by month’s end. Better reasoning on multi-step tasks, more reliable tool calling, improved comprehension of long documents, and cleaner structured outputs.

No configuration changes needed. Your existing agents just got smarter. If you’ve been building agents that chain multiple API calls or synthesize large SharePoint libraries, revisit your top prompts. You’ll likely notice improvements on the edge cases that used to require careful prompt engineering.

My Take on March

Three releases, one theme: declarative agents moved from “promising demo” to “production-ready stack.” Interactive UIs give agents real utility beyond text. The toolkit gives developers a clean, focused build experience. GPT-5.2 gives the model the reasoning chops for real workloads.

If you’re starting a new agent project today, the default architecture just shifted. MCP server with MCP Apps for the UI, embedded knowledge for domain context, GPT-5.2 handling the reasoning. That’s a capable foundation.