You set a completion condition on a substantial task and let it run without prompting each step, with a separate evaluator checking the condition after every turn. Partway through, the condition turns out to be impossible to satisfy. What happens?

The principle — a completion condition is not a finish line the worker runs at; it is a test applied from outside, repeatedly, and it has more than one way of coming true.

Why the key is correct — Anthropic gives the clearing rule in a single sentence: the goal clears automatically once the condition is met, if the model judges the condition impossible to satisfy, or if a turn fails on an error you have to fix. Impossibility is treated as a resolution, not as a reason to keep going. The evaluation that reaches that conclusion is the same separate evaluator that checks your condition after every turn, so completion is decided by a fresh model rather than the one doing the work — which is what makes the judgement worth anything. The setup around it is the ordinary one: you set a completion condition and Claude keeps working toward it without you prompting each step, and the guidance recommends this for substantial work with a verifiable end state.

Why the others are wrong — running forever is the intuition that second exit exists to defeat; an unsatisfiable condition is judged unsatisfiable and cleared. Clearing when the worker feels it has done enough puts the decision back where the arrangement took it from, and the whole point of a separate evaluator is that the party doing the work is not the party deciding it is finished. And evaluating once at the end would make this an exit check, which it is not: the condition is tested after every turn, which is how the run can stop of its own accord instead of having to be stopped and then assessed.

Remember this — a condition graded from outside, after every turn, ends the run in either direction: met, or found impossible.

Sources — Anthropic's Claude Code goal documentation.