An agent action is an operation selected or initiated by an AI agent toward a task, described with an explicit stage such as proposed, attempted, or observed as completed.
This operational definition keeps an action description separate from a claim that its effect occurred. The useful unit depends on the question: a resource update, a destination request, or a step in a larger task. State that unit before comparing action records or counts.
Choose a unit of work
One tool call can request several operations, and several calls can contribute to one task outcome. Retrying a request creates another attempt without necessarily changing the intended operation. A call, task, attempt, and effect are therefore not interchangeable units.
OpenTelemetry's traces documentation describes spans as units of work or operations and allows child spans for sub-operations. A span is a telemetry representation; its presence does not make it the business action or establish a one-to-one mapping to an external effect.
Make the execution stage explicit
Proposed: the system has selected or described an operation. The proposal can be evaluated, changed, denied, or left pending without the protected operation being attempted. A record of the proposal does not prove dispatch.
Attempted: an executor has begun the operation or dispatched a request at a specified boundary. Identify that boundary: sending a request to an intermediary is not the same event as the destination applying a change.
Observed outcome: a source reports a result or state relevant to the attempt. Preserve what that source establishes. A later state observation may show that a resource changed without identifying which attempt caused the change. A timeout may leave the outcome unknown.
The AI agent audit trail guide explains how to retain the identity, source, and causal evidence needed to relate those stages.
Permission and completion answer different questions
Authorization determines whether an operation may proceed under stated conditions. It does not report that the operation succeeded. A denial record likewise needs executor evidence before it can support a claim that the protected effect was withheld.
The pre-execution governance guide identifies where that effect can still be constrained. If work changes to a different target or operation, the relevant scope needs to be established for the changed proposal.
An action need not be expressed through a tool call in every agent implementation. The AI agent governance pillar connects the actual execution mechanisms to delegated authority, controls, and evidence.