How a loop knows to stop. Without an explicit answer it will either quit too early with a half-finished job, or keep going until it exhausts your budget. Deciding what ‘done’ looks like is one of the first things to specify and one of the most commonly left implicit.
Technical view
Conditions include the model signalling completion, a verifier passing, a maximum iteration count, a token or cost ceiling, a wall-clock timeout, and detection of no progress. Multiple conditions should be combined, with hard limits as a backstop independent of the model’s own judgement, since a model that believes it has finished is not a reliable stopping signal. Termination reason should be logged, as the distribution across reasons is a useful health measure.
Related
Part of AI Terminology Index, section: Agents and Tools.