>_ The Manifest

May was the month declarative agents got less magical and more useful. The biggest moves were not shiny demos. They were the kind of platform changes that make something survivable in a real tenant: evals, better MCP runtime plumbing, and government cloud support finally catching up to the interesting stuff.

Evals finally showed up

The public preview of the Microsoft 365 Copilot Agent Evaluations tool was easily the most important developer announcement of the month. Not the flashiest, just the one that matters if you want agents to graduate from “good demo” to “software.”

I keep coming back to the boringness of it. Good boring. Prompt sets, multi-turn conversations, judge-based scoring, HTML scorecards. That is exactly the stuff teams skip until the first regression sneaks into production and everybody realizes “I tried it twice” is not a test strategy.

What I like here is the framing. Microsoft is finally admitting that agent quality needs artifacts, not vibes. If you’re building declarative agents seriously, this should sit next to manifest validation and deployment checks. Same category. Same expectation.

A prompt-set entry looks like this in spirit:

[
  {
    "prompt": "What is our SLA for tier-1 incidents?",
    "expected_behavior": "Cites the SLA policy doc and quotes the response window.",
    "tags": ["sla", "policy-grounded"]
  }
]

Wire that into the same loop as manifest validation and pre-publish gates, and “I tried it twice” stops being your release criteria.

MCP stopped looking bolted on

The sharpest community post I found in May was this field report on team400.ai about adding skills to declarative agents. It makes a distinction I wish more people made: built-in capabilities are configuration, but API plugins and MCP tools are software. Different effort, different risk, different operational burden.

That post landed right as the roadmap started filling in the runtime story. Dynamic tool discovery for declarative agents means MCP tools can change without forcing a republish. Remote MCP server support through the Work IQ APIs means the platform is leaning into MCP as a real integration surface, not just a local dev convenience. And declarative agent access through the Work IQ endpoint, paired with the unified REST endpoint for agents and workflows, is the sort of plumbing you only build when you expect agents to be invoked as part of bigger systems.

That combination changed how I read the month. May did not add one giant new capability. It clarified the shape of the stack. Agents are becoming addressable runtime components with discoverable tools, remote execution surfaces, and actual lifecycle expectations. Good. That is much more interesting than another cute demo.

It also raises the standard. If MCP is your answer, you now need to think like a platform engineer. Auth, observability, rate limits, tool descriptions, versioning. The fun part is still there. The adult responsibilities are arriving right behind it.

Government cloud finally started getting the real story

The other signal worth paying attention to came from three roadmap entries on May 27: declarative agents with Actions in GCC High and DoD, MCP-based agent enablement for U.S. Government clouds, and support for interactive UI widgets in government clouds.

I like this because it closes a credibility gap. The pro-code story is not mature if all the interesting pieces only exist in commercial tenants. Actions matter. MCP matters. Interactive widgets matter. If government tenants are going to be asked to take declarative agents seriously, they need the same building blocks, not the watered-down version.

This is still roadmap language, so I would not confuse it with “done everywhere right now.” But the direction is clear. The platform is expanding the real agent surface area, not just the safe marketing layer.

My Take on May 2026

May was a plumbing month. I mean that as a compliment.

If you’re building right now, my advice is simple. Spend less time chasing one more capability flag. Spend more time on prompt sets, auth boundaries, tool contracts, and logs. That work got a lot less optional in May.

Resources

Have questions or want to share what you're building? Connect with me on LinkedIn or check out more on The Manifest.