A design document on operating long-running AI agents, covering harnesses, safety boundaries, verification, recovery, external memory, and human responsibility.