# What is Agent Run?

Also called execution, invocation.

An agent run is one bounded execution of an agent from trigger to termination, containing every step, tool call, and intermediate result produced along the way. It is the standard unit for measuring cost, duration, and success, and it is what logging, replay, and evaluation systems are usually built around.

A run has a clear anatomy: a trigger that started it, the inputs and state present at the start, an ordered sequence of steps, and a terminal outcome. Assigning each run an identifier and attaching that identifier to every model call, tool invocation, and log line is what makes it possible to reconstruct what happened after the fact.

Runs are distinct from sessions and from tasks. A session may span many runs, as when a user sends several messages over an afternoon. A task may require several runs, as when a scheduled job retries after a failure. Metrics that mix these levels are misleading, since cost per run and cost per task can differ by an order of magnitude.

Terminal outcomes should be recorded with more granularity than success or failure. Completed, stopped at a budget, blocked by a guard, escalated and awaiting a person, and errored are genuinely different endings that call for different responses. Collapsing them hides the most actionable information about how a system behaves in production.

The run is also the natural granularity for cost and quality attribution. Tokens, tool calls, wall clock time, and outcome all attach to it cleanly, which makes it possible to ask which triggers produce expensive runs and which produce useless ones. Without a run identifier threaded through the system, those questions can only be answered in aggregate.

## Key points

- One bounded execution from trigger to termination
- Run identifier threads through every call and log line
- Distinct from a session and from a task
- Record granular terminal outcomes, not just pass or fail
- Natural unit for cost, latency, and quality attribution

## In practice

A nightly schedule triggers a reconciliation agent, producing run 8842. It makes 31 tool calls, 12 model calls, and finishes in four minutes with the outcome completed. The next night run 8843 is triggered identically but ends as stopped at budget after 40 steps. Comparing the two trajectories by run identifier shows a new vendor format sending the second run into repeated parse retries.

## Related terms

- [Agent Trajectory](/en/glossary/agent-trajectory)
- [Agent Loop](/en/glossary/agent-loop)
- [Termination Condition](/en/glossary/termination-condition)
- [Step Budget](/en/glossary/step-budget)
- [Checkpointing](/en/glossary/checkpointing)

[Back to the AI Glossary](/en/glossary)