Your AI Agent Needs a Maintenance Window Protocol

A maintenance window protocol is proposed for long-running AI agents to handle deployment, credential rotation, and host restarts. The protocol has four goals: stop new work, let safe work finish, make ambiguous work visible, and resume with an explicit recovery decision. The protocol uses a durable state machine with states RUNNING, DRAINING, QUIESCED, STOPPED, and NEEDS_REVIEW. The state is persisted to disk, and checkpoints are recorded before and after non-idempotent operations. This allows for safe recovery and reconciliation of work.