A tool call is a request to perform an operation through a tool interface, identifying the tool or operation and supplying the inputs that the interface requires.
This operational definition describes the request. To interpret an event or diagram, also establish its stage: a proposed call may not have been dispatched, an invocation may be awaiting a response, and a completed response may leave the external effect unconfirmed. The words “tool call” alone do not distinguish those states.
What the request contains
The interface determines the operation identifier and required inputs, often called arguments. These can select a resource, supply material to process, or specify a change. The same tool name with different inputs can describe operations with different authorization requirements.
MCP's Tools specification represents a call using a tool name and arguments and distinguishes protocol errors from tool-reported errors. That is one protocol's representation. A tool call does not require MCP, a network request, or the same input format in every agent implementation.
A proposed call is not proof of an effect
An AI agent can select an operation as part of pursuing a task. If the proposed request is held for policy evaluation, its presence in the workflow does not establish that the protected operation occurred. Even when a tool service receives the request, a later action at the destination may still be pending.
Hypothetical example: an agent proposes retrieving a document's metadata. A record of the proposal does not prove dispatch. A tool response reporting an access failure does not prove that the requested metadata was returned. The meaning of each observation depends on where it was produced and what it records.
Invocation, result, and action have different scopes
A tool can perform work locally or delegate to another component. One invocation may lead to several operations, and a task can require several calls. Do not assume a one-to-one relationship between a call and an external effect.
An action attempt and its result need evidence appropriate to the executor and destination. A timeout can leave the outcome unresolved; a permission decision authorizes an attempt without guaranteeing completion. Conversely, not every agent action must be expressed as a tool call.
The AI agent governance architecture guide explains how to connect proposed calls, policy decisions, execution, and evidence without treating the proposal as the completed action.