We run agents in production. For months our biggest reliability problem wasn’t the models being wrong. It was the models saying “done” when nothing happened. The patterns, in order of how often we saw them: The agent reports success and the tool actually…