This Month In Declarative Agents: JSON Is the New C#

February was a quiet month for declarative agents content. Only two items passed the filter. But sometimes a quiet month is the most telling kind: when the volume drops, the signal gets clearer.

The Afternoon Agent

A walkthrough on rickvanrousselt.com builds a “SharePoint Site Advisor” declarative agent from scratch using the M365 Agents Toolkit. That part isn’t unusual. People are building agents all the time. What makes this post worth your time is the framing.

The post opens by describing what it took to build a conversational bot in 2017: weeks of C#, Bot Framework plumbing, LUIS model training, and a painful amount of dialog state management. Then it flips to today, where the equivalent functionality lives in a JSON manifest, a plugin definition, and an afternoon of configuration.

That contrast isn’t nostalgia. It’s the whole point. Declarative agents have compressed what used to be a full-stack engineering project into a configuration exercise. And the post is honest about where the boundary still sits. If your scenario needs custom middleware, real-time event handling, or anything beyond what the Copilot orchestrator provides, you still need a proper bot. The JSON-first approach doesn’t replace everything. It replaces enough that most internal tools never need to leave it.

I’ve been saying for a while that the developer story for declarative agents is underrated. Not the “build your first agent” tutorial kind of developer story. The “wait, this used to take three sprints and now it doesn’t?” kind. This post captures that shift better than most documentation pages I’ve read. You don’t need to sell people on AI anymore. You need to show them that the thing they spent half a quarter building five years ago now takes a lunch break.

Scanned PDFs Join the Grounding Party

On the platform side, Microsoft quietly shipped a capability that matters more than the headline suggests. Declarative agents can now ground answers in scanned PDFs and image-based documents stored in SharePoint.

This is a big deal for anyone building agents against real enterprise content. If you’ve worked in regulated industries or large organizations, you know the pattern: a huge chunk of SharePoint content is scanned contracts, signed forms, faxed approvals, and legacy documentation that was never born digital. Until now, declarative agents simply couldn’t see any of it. The files were there, SharePoint indexed them, but the agent’s grounding pipeline skipped right over them.

That gap silently undermined trust. Users would ask a question, the agent would miss the answer because the source document was a scanned PDF, and nobody would understand why the agent seemed unreliable. It wasn’t a bug anyone would file. It was just a quiet failure that eroded confidence over time.

Now that gap is closed. If you’re building a knowledge agent for legal, compliance, HR, or any team that deals with physical-origin documents, go test this immediately. It changes what your agent can reliably answer, and it removes one of those invisible failure modes that made early deployments feel flaky.

My Take on February

February was thin on volume but sharp on signal. One community post reframed the developer experience narrative better than most official content could. One platform update removed a grounding blind spot that’s been quietly tripping up enterprise deployments.

Not every month needs a dozen posts to matter. Sometimes the story is that the ecosystem is settling into something usable, something predictable. Declarative agents aren’t the flashy new thing anymore. The tooling works. The platform is filling gaps. The community is writing about real implementations instead of experiments.

That’s where we are right now, and honestly, that’s a good place to be.