Why AI Agents Need a Formal Maintenance Protocol Before Shutdown

Long-running AI agents face a critical gap when deployments, credential rotations, or host restarts occur mid-operation, as standard process supervisors cannot determine whether in-flight tool calls were safely completed. A proposed maintenance-window protocol guides agents through four states — RUNNING, DRAINING, QUIESCED, and NEEDS_REVIEW — to ensure work is either finished, checkpointed, or flagged for review before shutdown. The protocol distinguishes between replayable LLM steps and non-idempotent actions such as payments, emails, or Git pushes, requiring checkpoints immediately before and after each irreversible operation. Idempotency keys tied to logical operations rather than random process attempts allow safe reconciliation on restart. The approach also mandates durable state persistence, drain timeouts, and staging-environment testing across key interruption points to validate that the protocol behaves correctly under real failure conditions.