Durable AI Agents With Leased Workers Could Scale to Millions Without Bloat

A widely used approach to building AI agents treats them as long-running processes, but this wastes resources and makes crashes costly when tasks stretch over minutes or hours. A proposed alternative separates the agent's state from the worker executing it, storing progress externally so any temporary worker can pick up where another left off. Workers would operate on short time leases, handling several steps before checkpointing and exiting, rather than idling while waiting for external events like CI builds or human approvals. This architecture could allow millions of concurrent agent runs without requiring millions of active processes, since compute is only consumed when an agent has actionable work. The approach mirrors how modern web systems handle user sessions and introduces known infrastructure challenges around state reconstruction, idempotent side effects, and independent streaming.