StopLine

StopLine stops a tree of AI agents mid-run and accounts for every action it had already started. A stop flag fails in a simple way: a worker reads the flag, then acts after someone pressed stop, and nothing records what was already on its way. StopLine puts one check in MongoDB Atlas in front of every agent action. Permission is data: who granted it, to whom, for which tools and targets, with what limits. A repair agent hands narrower permissions to its workers, and each action is checked and recorded in one transaction. Stop revokes a permission and everything handed down from it, cancels work that was allowed but not yet sent, and writes a receipt listing every action already started. It then asks each target system about anything that was sent and has no answer: if it never arrived it is fenced there so a late send is refused, and if the target cannot be reached it stays "unknown". The receipt closes only on evidence from the target. The agents act on a real Kubernetes cluster and a traffic router. Measured over 43 stop schedules, a stop flag let an action start after the stop in 20; StopLine in 0. A live Deep Agents repair agent runs through the same check, a reviewer's report rebuilds any run from the stored records, and a reviewer can ask the audit trail questions in plain words.