How to Prevent AI Agents from Losing or Duplicating Work Under Queue Pressure

Long-running AI agents frequently fail not due to model errors but because of poorly managed transitions between work states such as queued, leased, running, and completed. A relational database schema with explicit job statuses and a unique idempotency key can make these states visible and trackable. Engineers are advised to define concurrency limits per tenant and worker pool, and to persist the queue before acknowledging jobs to avoid data loss on process crashes. Workers should claim jobs atomically with short leases and verify lease ownership before each externally visible step, since expired leases can cause two workers to execute the same job simultaneously. For operations with external side effects like emails or deployments, recording intent before execution and reconciling unknown outcomes via provider APIs is essential to avoid duplicate actions.