Agent Tool.

An agent tool is a capability made available to an AI agent so it can do something beyond generating a response. A tool might retrieve information, query a system, run a calculation, or submit a defined request to another service. The agent decides whether the tool is relevant; the tool performs the operation it was designed and authorized to perform.

The term is not a single, universal Microsoft product name. Depending on the platform, a similar capability may be described using terms such as an action, connector, function, or integration. These labels can have specific meanings in their respective products, so they should not be assumed to be interchangeable.

An agent and a tool have different responsibilities. The agent interprets the user’s request, determines whether it needs additional information or an action, and uses the tool’s result to continue. The tool provides a defined operation, such as looking up a record or submitting a request. It may be ordinary software rather than another AI model.

This separation helps make an agent’s capabilities easier to reason about. If an agent can only call an approved lookup tool, for example, it should not be treated as having general access to every system connected to the organization. What it can actually do depends on the tools made available, their configuration, and the identities and permissions used when they run.

A tool interaction generally follows a short cycle:

1. Interpret the request: The agent identifies information it needs or an action it has been allowed to take.
2. Prepare the call: It supplies the tool with required inputs, such as a record identifier or search terms.
3. Run the operation: The tool performs its configured task using its available permissions.
4. Inspect the response: The agent uses the returned result, asks for clarification, or reports that it cannot proceed.

A well-designed tool has a narrow purpose and clear inputs and outputs. The agent’s ability to decide when to call it does not replace validation of those inputs, enforcement of permissions, or review of actions that have significant consequences.

Tools may support different kinds of work:

• Information retrieval: Find a record, search an approved knowledge source, or check a status.
• Calculation or transformation: Apply a defined operation to supplied information.
• Workflow initiation: Submit a request or start a process subject to configured controls.
• System integration: Exchange information with an approved business application or service.

A tool can return information without changing a system, or it can initiate an action. That distinction is important when evaluating risk. A lookup that displays a ticket is not equivalent to a tool that changes its assignment or closes it.

Suppose an employee asks an assistant for the status of a support request. The agent may recognize that it needs current ticket information and call a ticket-lookup tool with the request identifier. The tool returns the fields it is configured to expose, and the agent summarizes the result for the employee.

The tool does not automatically have permission to make every change to that ticket. If the workflow also allows reassignment or closure, those should be separate, clearly defined operations with their own access controls. This makes it easier to limit what the assistant can do and to investigate what happened if a request produces an unexpected result.

Microsoft platforms can provide ways to connect agents with data, business applications, or actions, but the available mechanisms and terminology depend on the product and configuration. An agent-building experience, such as Copilot Studio, may present tool-like capabilities differently from developer-focused AI environments. Confirm the platform’s terminology and behavior rather than assuming that a feature named “tool” works identically across products.

An agent tool is also distinct from Copilot Chat and Microsoft 365 Copilot. Copilot Chat is a conversational AI experience whose information access depends on account type, enabled capabilities, and configuration. Microsoft 365 Copilot is designed to connect AI assistance more directly with Microsoft 365 work contexts and applications, subject to user permissions and organizational controls. A tool is an operation an agent may use within an experience or workflow, not a separate name for either Copilot experience. Licensing and eligibility may vary by subscription, application, account type, tenant configuration, region, and service update.

A tool’s access should match its purpose. If it only needs to read a limited set of records, it should not be configured with broader access without a specific reason. Organizations should also determine whose identity and permissions apply when the tool runs, since that can affect which information is available and which actions are allowed.

Tool calls can fail, return incomplete information, or produce results that do not answer the user’s question. Treat returned content as data to evaluate, not automatically as a trusted instruction. For operations that change records, send messages, or affect access, define approval and confirmation requirements, retain appropriate logs, and provide a safe way to stop or escalate the workflow.

An agent tool is a bounded capability that lets an agent retrieve information or carry out a defined operation. The agent handles interpretation and flow; the tool handles the configured task. Clear boundaries between the two help teams understand what the agent can access, what it can change, and where a person should remain involved.

In Microsoft environments, verify the exact product, tool behavior, identity model, and permissions in use. The label alone does not establish what data is accessible or which actions are possible.