Building Agents on M365 Copilot¶

Part of: Build Workflows > Agents

Microsoft 365 Copilot extends its built-in capabilities through agents — specialized AI assistants tailored to specific domains. Agents can retrieve information, summarize data, and take actions like sending emails or updating records, all within the Microsoft 365 apps where you already work.

Official docs: Agents for Microsoft 365 Copilot

There are two approaches to building agents, depending on how much control you need.

Agent Core Components¶

All M365 Copilot agents share these building blocks:

• Knowledge — specialized instructions and data sources that shape the agent's responses
• Actions — triggers and workflows that automate business processes
• Orchestrator — the engine that manages how the agent interacts with knowledge and actions
• Foundation models — the AI models powering reasoning and language understanding
• User experience layer — how users interact with the agent across Microsoft 365 apps

Declarative Agents¶

Declarative agents use Copilot's built-in AI infrastructure, model, and orchestrator. You configure them by adding custom instructions, knowledge, and actions — no hosting required. Because they run on Copilot's infrastructure, they inherit Microsoft 365 security, compliance, and Responsible AI standards.

What you configure¶

• Custom instructions — shape Copilot's responses for your organization's workflows
• Custom knowledge — connect Microsoft 365 data sources (Teams, SharePoint, OneDrive) or external data via Copilot connectors
• Custom actions — integrate with APIs to interact with external systems in real-time

How to build them¶

Where they run¶

Declarative agents run within Microsoft 365 Copilot and Microsoft 365 apps — Teams, Word, Excel, Outlook, and SharePoint. Users invoke them via @mentions or in business chats.

When to use declarative agents¶

• You want agents within Copilot's orchestration with built-in security and compliance
• Your users work within Microsoft 365 apps and want agents in that context
• You want faster implementation with no-code or low-code tools
• You don't need custom AI models or external hosting

Mapping your Design blueprint¶

Example: Implementing your Design blueprint (Copilot Studio)¶

Here's how a research analyst agent from the Design phase maps to Copilot Studio's visual builder:

1. Name → Enter Research Analyst as the agent name when creating a new declarative agent
2. Description → Add the agent's scenario description: "Researches companies and produces structured briefs for the team"
3. Instructions → In the Instructions section, paste your agent instructions: "You are a research analyst who investigates companies and produces structured briefs. When asked about a company, search for information and summarize findings with sections for: company overview, key products/services, recent news, and relevant insights."
4. Model → Copilot's built-in foundation models are used automatically — no model selection needed
5. Tools → In the Actions section, add connectors: enable Web Search for research, add SharePoint knowledge sources for internal data, or configure API plugins for external system access

Example: Implementing your Design blueprint (JSON manifest)¶

For developers using the Microsoft 365 Agents Toolkit, the same agent is defined as a declarative JSON manifest:

{
  "$schema": "https://developer.microsoft.com/json-schemas/copilot/declarative-agent/v1.5/schema.json",
  "name": "Research Analyst",
  "description": "Researches companies and produces structured briefs",
  "instructions": "You are a research analyst who investigates companies and produces structured briefs. When asked about a company, search for information and summarize findings with sections for: company overview, key products/services, recent news, and relevant insights.",
  "capabilities": [
    { "name": "WebSearch" }
  ],
  "conversation_starters": [
    { "text": "Research Acme Corp and write a brief" },
    { "text": "What can you tell me about Contoso Ltd?" }
  ]
}

To add knowledge sources (SharePoint, OneDrive) or API actions, add them to the manifest's capabilities and actions arrays. See the Agents Toolkit documentation for the full schema.

Custom Engine Agents¶

Custom engine agents are fully customized AI assistants. You bring your own orchestrator, choose your own models, and host the agent outside of Microsoft 365. This gives you full control over workflows, business logic, and integrations — but you're responsible for compliance, security, and hosting.

What you configure¶

• Custom orchestration — full control over workflows, with one or more language models
• Custom models — choose any model (foundation, fine-tuned, or industry-specific)
• Autonomy — agents can initiate workflows proactively, make decisions, and escalate tasks

How to build them¶

Where they run¶

Custom engine agents run in Microsoft 365 apps (Teams, Word, Outlook) and external apps and websites — customer service portals, internal dashboards, or any web application.

When to use custom engine agents¶

• You need custom orchestration for complex workflows or specific business logic
• You want to use your own AI models (domain-specific, fine-tuned, or multimodal)
• You need group collaboration where multiple users interact with the same agent
• You want agents available outside Microsoft 365
• You need proactive messaging — agents that trigger actions without user input
• You're integrating an existing conversational assistant with Microsoft 365

Mapping your Design blueprint¶

For multi-agent workflows, custom engine agents support agent-to-agent communication — agents can delegate tasks and coordinate workflows across your organization.

Choosing Between Approaches¶

Start with declarative agents if your workflow fits within Microsoft 365 and you want built-in security and faster setup. Move to custom engine agents when you need custom models, external integrations, or proactive behavior.

What's Next¶

• Agents overview — the platform-agnostic agent decision framework
• Design Your AI Workflow — produce the agent blueprint that feeds into these implementations