Microsoft 365 Copilot Agents: Examples and How to Build One
Microsoft 365 Copilot agents are specialized assistants inside Copilot, from Microsoft’s own Researcher and Analyst to ones you build yourself. How they work, real examples, building one in Agent Builder, and adding an MCP server to a Copilot Studio agent so it can act in other systems.
7 min read
Microsoft 365 Copilot agents are specialized assistants that run inside Copilot Chat, Teams, Outlook and the other Microsoft 365 apps. Each one adds instructions, knowledge and sometimes actions for a particular job. Microsoft ships some, such as Researcher for deep, cited research and Analyst for data work, and you can build your own two ways: in Agent Builder, inside the Microsoft 365 Copilot app, for a quick question-and-answer agent over your content, or in Copilot Studio when the agent needs to act in other systems. The fastest way to let a Copilot agent act outside Microsoft 365 is to add a remote MCP server to a Copilot Studio agent, which the second half of this post walks through.
How Microsoft Copilot agents work
Microsoft’s overview of agents for Microsoft 365 Copilot (opens in a new tab) describes two core components, knowledge and actions, driven by an orchestrator and foundation models. There are two kinds of agent:
- Declarative agents use Copilot’s own orchestrator and models. You supply custom instructions, knowledge such as SharePoint, OneDrive, Teams messages and Copilot connectors, and optionally actions that call APIs. They need no hosting of their own and inherit Microsoft 365’s security and compliance. Agent Builder and the Microsoft 365 Agents Toolkit both make them.
- Custom engine agents bring their own orchestration and models. They can start work without a user prompt and run outside Microsoft 365, but you host them and answer for their compliance. Copilot Studio or pro-code tools build them.
Either way, an agent works with the permissions of the person using it. Microsoft’s governance notes for Agent Builder put it as “no new privileges”: if you cannot open a SharePoint site or a mailbox, the agent will not show you content from it.
Examples of Microsoft Copilot agents
Start with the agents Microsoft already provides, which you find under Agents in Copilot Chat or in the Agent Store:
- Researcher: according to Microsoft’s page on the Researcher agent (opens in a new tab), it handles complex, multi-step research across the web and your files, emails, meetings and chats, and returns a structured report with cited sources. It takes longer than standard Copilot chat by design.
- Analyst: Microsoft’s Researcher FAQ points Excel and spreadsheet work to Analyst instead of Researcher.
- Planner Agent: works tasks in Planner and writes status reports, covered in Microsoft Copilot for project management.
For agents you build, Microsoft’s own examples split by tool. Agent Builder suits a project FAQ agent that answers from the project’s documents, a product documentation assistant over internal manuals, or an onboarding agent for new team members. Copilot Studio suits a support agent that creates tickets and escalates to a person, an IT help desk triage agent, or a sales assistant that reads CRM data and starts an approval.
Build an agent in Agent Builder
Microsoft’s guide to building agents with Agent Builder (opens in a new tab) starts from New agent in Microsoft 365 Copilot, on the web or in the Teams desktop and web clients. You can describe the agent in plain language and let Copilot fill in the configuration, start from a template, or choose Skip to configure and set it up by hand:
- Name: up to 30 characters, descriptive and unique.
- Description: up to 1,000 characters. Copilot uses it to decide when your agent fits a request, so keep it short and precise.
- Instructions: up to 8,000 characters on what the agent does and how.
- Knowledge: public websites, SharePoint content, embedded files and Copilot connectors. With a Microsoft 365 Copilot license, users can also ground it in their Teams chats and Outlook email.
- Starter prompts, so others see what the agent is for.
- Test on the Try it tab, then create the agent and share it.
Code interpreter and image generation are available as capabilities. The instructions do most of the work, so write them like a brief for a new colleague:
You answer questions about the Website Relaunch project for the marketing team. Use only the project SharePoint site and the files attached here. Answer in plain English, in five sentences or fewer, and cite the file. Dates are US style (October 14, 2026). If the sources do not answer the question, say so and suggest who to ask. Never guess a date, a cost or an owner.
Agent Builder has no actions. When an agent needs to call an outside service, copy it to Copilot Studio, which keeps its configuration and instructions. What Copilot Studio adds, and how it is licensed, is in Microsoft Copilot Studio: what it is and when to use it.
Add an MCP server to a Copilot Studio agent
MCP lets an agent use another system’s tools without a custom connector. Microsoft’s guide to connecting an agent to an existing MCP server (opens in a new tab) recommends its MCP onboarding wizard. The server must be remote and speak the Streamable HTTP transport, and generative orchestration must be on:
- Open the agent in Copilot Studio, go to its Tools page and select Add a tool, then New tool, then Model Context Protocol.
- Fill in Server name, Server description and Server URL. The orchestrator reads the description to decide when to call the server, so say plainly what it is for.
- Choose authentication: None, API key, or OAuth 2.0. For OAuth, Dynamic discovery is the simplest when the server supports dynamic client registration with discovery; otherwise use Dynamic or Manual and enter the endpoints.
- Select Create, then Create a new connection, sign in, and select Add to agent.
- Turn off any of the server’s tools the agent does not need, test it in the test panel, then publish and connect the Teams and Microsoft 365 Copilot channel.
Some servers are in the list already. Box, for example, appears under Model Context Protocol once a Box admin makes the integration available. If an option is missing, ask your admin: data policies that govern Power Platform connectors govern MCP servers too.
The pro-code route
Developers who build declarative agents in VS Code can wrap an MCP server as a plugin instead. Microsoft’s walkthrough on building a plugin from an MCP server (opens in a new tab) uses Microsoft 365 Agents Toolkit 6.12.0 or later: create a Declarative Agent, choose Add an Action, then Start with an MCP Server, and pick OAuth with static or dynamic registration, Entra SSO, or none. The toolkit discovers the server’s tools at run time.
Microsoft Copilot agent best practices
- One job per agent. “Answers questions about the relaunch” beats “helps marketing.”
- Write the description for the orchestrator. It decides when your agent, or your MCP server, gets called.
- Keep user credentials. A Copilot Studio tool set to maker-provided credentials lets everyone act as the maker.
- Read before write. Add only the tools the job needs, and keep anything that sends, deletes or changes records behind a person.
- Test on yourself first. Microsoft advises publishing for your own use in Teams and testing before making an agent available to others.
- Instruct it never to fill gaps. An agent that says “not in the sources” is worth more than one that guesses.
Giving a Copilot agent a task list to work from
Agents in Teams are often asked to “add that to the list.” fenbs is a task board where people and AI assistants are members with roles, and its MCP server at https://fenbs.ai/api/mcp supports the Dynamic discovery option, so each user signs in to fenbs and chooses what the agent may do. Tell the agent in its instructions to call fenbs_whoami and fenbs_get_context first: the context starts with the rules on the Decisions and rules page, then the AI context notes every assistant reads. It then files features, enhancements and bugs with a note, a plan and a priority from 1 to 10, where 1 is the most urgent, and History records each change with who made it. fenbs has no due dates, no settable assignee and no sprints, so ask the agent to put any deadline or owner in the task note. More on writing tasks an assistant can finish is in how to write a task for an AI agent.
Related
What Copilot Studio is and how it is licensed: Microsoft Copilot Studio. Mail and calendar tools for the same agents: Outlook MCP server. Keeping the risky steps with a person: AI agent approval workflow. Connecting fenbs: MCP docs.