Declarative Agents for Microsoft 365 Copilot
Declarative agents specialize Microsoft 365 Copilot through a JSON manifest — instructions, knowledge, and OpenAPI actions — running entirely on Microsoft's orchestrator with no hosting of your own. We examine the schema as of August 2025, governance in the admin center, the metering economics, and the concrete cases where pro-code custom engine agents remain the better choice.
Specialized assistants without new infrastructure
Every department wants its own assistant: one that knows the travel policy, the ticket backlog, the product wiki. The default engineering answer in 2025 is a RAG stack — embeddings, a vector store, retrieval code, authentication, hosting, monitoring, and a compliance review before anyone types a first prompt. For an assistant that answers questions about documents that already sit in SharePoint, that is a lot of machinery to own and operate.
Microsoft's alternative is the declarative agent, generally available since November 2024. You do not build a runtime; you describe one. The description runs on the same orchestrator, foundation models, and trusted AI services that power Microsoft 365 Copilot itself. This article examines what that buys, what it costs, and where it stops.
Anatomy of a declarative agent
A declarative agent is a JSON manifest inside a Microsoft 365 app package. It has three functional parts: instructions that shape behavior, capabilities that declare knowledge sources, and actions that reference API plugins, plus up to 12 conversation starters. There is no code and no infrastructure of your own — no bot endpoint, no model deployment, no vector database. Microsoft hosts, orchestrates, and updates everything.
Authoring routes match skill levels: Agent Builder inside Copilot for makers, Copilot Studio for low-code teams, and the Microsoft 365 Agents Toolkit in Visual Studio Code — renamed from Teams Toolkit at Build in May 2025 — for developers who keep manifests in source control. Identity comes from Microsoft Entra; data access is bounded by the signed-in user's existing permissions.
Instructions and knowledge
Instructions are a required manifest field, capped at 8,000 characters. They function as a system prompt for the Copilot orchestrator: role, workflows, tone, restrictions. Microsoft's guidance is explicit that instructions belong in the manifest, not in referenced documents. Knowledge sources are grounding data, not a trusted instruction channel — directive-like content there can be sanitized or ignored at runtime.
Knowledge is declared through capabilities, and the list grows monthly. Schema 1.2 added scoped web search — at most four sites — plus code interpreter and image generation. Version 1.3 (March 2025) added Dataverse, Teams messages, and people knowledge. Version 1.5 (July 2025) added meetings, bringing the total to eleven capability types. SharePoint, OneDrive, and Copilot connector sources can be scoped to specific sites, folders, or connections.
Actions through OpenAPI plugins
Actions connect the agent to REST APIs. Each action is an API plugin: an OpenAPI description plus a plugin manifest that adds semantics, authentication, and confirmation behavior. The Copilot orchestrator decides when to call which operation; consequential operations trigger a user confirmation by default. A precise OpenAPI document is not optional here — it is the planner's only view of your API.
Be clear about the boundary. Actions are request-response. There are no long-running jobs, no background triggers, no custom planning loop. If tool selection fails, you tune descriptions and instructions — you cannot replace the planner. That is the trade: you give up orchestration control and get Microsoft's operational surface without running any of it.
Governance in the Microsoft 365 admin center
Governance is where the model earns its keep. The Copilot Control System in the Microsoft 365 admin center gives administrators an inventory of agents across Copilot Studio, Agent Builder, and the store. Employee-built agents require admin approval before tenant-wide publication: administrators review data sources, permissions, and actions, then publish to selected users or groups — or block and reject.
The May 2025 update added agent usage reports, message consumption reports, and pay-as-you-go billing setup directly in the admin center. Microsoft Purview extends data security posture management, audit logging, and eDiscovery to agent interactions. None of this is code you write. It is tenant configuration — which is exactly why the declarative model appeals to regulated organizations.
Licensing and metering
Two access paths have existed since January 2025. Users with a Microsoft 365 Copilot license use declarative agents within fair-use limits at no extra charge. Users with the free Microsoft 365 Copilot Chat reach agents through metered billing: responses consume messages, and grounding in tenant data is metered at a substantially higher rate than plain generative answers. Model the cost before rollout, not after.
| Access option | Price (USD as of Aug 2025) | Billing |
|---|---|---|
| Microsoft 365 Copilot license | $30 per user per month | Agent use included within fair-use limits |
| Prepaid message pack | $200 per tenant per month | 25,000 messages, purchased in the admin center |
| Pay-as-you-go | $0.01 per message | Billed in arrears through an Azure subscription |
Where pro-code wins
Declarative agents stop where the platform's assumptions stop. You do not choose the model. Behavior shifts when Microsoft rotates Copilot's underlying models, so instructions need periodic re-testing. Output surfaces are Microsoft 365 Copilot experiences only — no public website, no external channel. Retrieval ranking, chunking, latency, and evaluation tooling are not yours to change.
For those cases Microsoft's own taxonomy points to custom engine agents: your orchestrator, your models, built with the Microsoft 365 Agents SDK (announced at Ignite 2024) or Azure AI Foundry, hosted and paid for by you. We draw the line pragmatically: declarative for Q&A and light actions inside the Microsoft 365 boundary; pro-code when you need deterministic workflows, pinned model versions, custom evaluation, external channels, or latency guarantees.
Outlook from August 2025
As of August 2025 the manifest schema has shipped three revisions in seven months. We expect the cadence to hold and two lines to converge. First, actions: Copilot Studio gained Model Context Protocol support in 2025, and it would surprise us if declarative actions did not follow — MCP servers as the standard backend behind API plugins.
Second, models: the scenario-models capability and the Copilot Tuning preview announced at Build 2025 suggest that 'no model choice' is softening into 'curated model choice'. If both land, the manifest becomes less a limitation and more a governance artifact: a declared, reviewable contract over increasingly capable runtimes. That is worth building toward — deliberately, and with the metering dashboard open.
Sources
- Copilot is the UI for AI: Introducing Microsoft 365 Copilot Chat — Microsoft 365 Blog (15 Jan 2025)
- Agent management updates in the Copilot Control System — Microsoft Mechanics / Tech Community (19 May 2025)
- Take Control of Agent Costs — Microsoft Power Platform Blog (21 Jul 2025)
- Declarative agent schema 1.5 for Microsoft 365 Copilot — Microsoft Learn (Jul 2025)
- Write effective instructions for declarative agents — Microsoft Learn (2025)