What it does

A declarative agent runs on Microsoft 365 Copilot's own model and orchestrator. You supply instructions, knowledge sources, and actions; Copilot does the reasoning. It's the fastest way to ship a focused agent.

Key facts

• Built with Agent Builder (no-code), Copilot Studio, or Teams Toolkit and the agent manifest.
• Knowledge can point at SharePoint/OneDrive, Graph connectors, web content, and embedded files.
• Actions come from API plugins (OpenAPI), connectors, or MCP servers.
• Instruction text has a character ceiling (thousands of characters, not unlimited) — keep prompts tight.
• Needs a Microsoft 365 Copilot licence to run; packaged and distributed as an M365 app.

When to use / skip

Use for the vast majority of "Copilot that knows about X and can do Y" asks. Skip only when you need a different model, custom orchestration, or a fully standalone chat surface outside Copilot.

Configuration decisions

• Which knowledge sources, and whether they're already indexed.
• Instruction scope — how tightly you constrain behaviour.
• Which actions to expose, and their auth.

Gotchas

• Grounding is only as good as the underlying index; unindexed SharePoint content simply won't surface.
• Users still need access rights to the source content — the agent doesn't bypass permissions, and it shouldn't.
• Instruction length limits bite once you try to encode complex logic.

Consultant notes

• This is the default recommendation. Cheap to build, easy to govern, no model hosting.
• Grounding cost sits with the connectors and Copilot licensing, not the agent itself — but connector item counts add up.
• Approval to publish tenant-wide goes through the admin centre, so agree the sign-off route early.

---

Review if the manifest schema or instruction limits change.