Introduces durable loops for AI agents built on Conductor, explaining why agent systems are fundamentally loops, why traditional in-process loops fail as tasks grow longer, and demonstrating a self-refining agent using the DO_WHILE task that survives crashes.