What is a Copilot agent?

A Copilot agent is a task-focused assistant inside the Microsoft Copilot ecosystem. Where Copilot chat answers a question, an agent is scoped to a job: it carries its own instructions, draws on a defined set of knowledge, and often follows a repeatable set of steps. If you want the ranked shortlist instead of the definition, go to the best Copilot agents.

The plain definition

Microsoft describes agents as specialized assistants that go beyond a general chat: they can answer questions, automate tasks, and run workflows, using the tools and data they are given. The important word is scoped. An agent does not try to be good at everything. It is pointed at a task and given the context to do that task well.

That scoping is what makes an agent more reliable than a clever prompt. A prompt is a one-time instruction. An agent packages the instruction, the sources, and often the steps together, so the same task produces a consistent result next week as well as today. For Microsoft's own wording, see its Agents for Microsoft 365 Copilot overview.

Copilot chat vs a Copilot agent

The main types of Copilot agent

Microsoft groups agents into two build types, plus the ready-made agents that ship inside Copilot. The distinction matters because it tells you how much work an agent takes and who can build it.

Declarative agents

A declarative agent is defined by what you tell it: a name and description, a set of instructions, and the knowledge it may use. There is no custom code to host. You configure it in Microsoft's agent-building tools. It is the right shape for a team process that can be written down — onboarding answers, a review checklist, a scoped assistant for one department.

Custom engine agents

A custom engine agent is software. Developers build it with their own runtime, code, or model, and connect it into the Copilot experience. It is the right shape when the task needs logic, integrations, or a model that Microsoft's declarative agents do not cover. It costs real engineering time.

Built-in agents

Some agents are ready to use without building anything. Two of the best-documented are Researcher, for multi-step research across web and work content, and Analyst, for working with data and numbers. Others cover prompting, long documents, and meetings. These are where most people should start, because they need no build.

What an agent is made of

Whatever the type, the useful parts are the same. Knowing them helps you tell a good agent from a thin one:

• Instructions: what the agent is for, how it should behave, and what it must not do.
• Knowledge: the sources it may draw on — files, sites, and connected systems — limited to what the user is allowed to see.
• Tools and actions: optional steps it can take, such as retrieving a record or drafting a message.
• A trigger or entry point: how you reach it — a chat, a channel, a document, or an app.

What an agent is not

• Not a general chatbot. The value is in the narrowing, not in knowing everything.
• Not always-on and unsupervised. It works within the permissions and limits set for it.
• Not a source of truth. It summarizes and reasons; a person still verifies anything that matters.
• Not free by default. Availability depends on plan and licensing.

When you need an agent, and when a prompt will do

You probably need an agent when the task repeats, when the same context is needed every time, when several people should get the same treatment, and when the sources are fixed. A prompt is enough when the task is one-off, when you are exploring, or when the request changes every time. Most people start with prompts and graduate to an agent once a task becomes routine.

Once you have a task in mind, the practical next step is to see how people actually use these tools day to day. Our guide to Copilot agents for work is organized by job to be done.

TopCopilotAgents is independent and not affiliated with Microsoft. General information only; product names and behavior change, so confirm details against Microsoft's own documentation.