Autonomy, with an exit. Give every AI-agent run a hard boundary for time, tools, retries and cost—then see exactly why it stopped.