Once an agent runs longer than a normal request-response interaction, the engineering problem changes. The process can crash halfway through. The context window can fill after many tool calls. The user can close the browser and come back later expecting progress. A worker can be evicted. A model call can time out. A bad loop can keep spending money long after anyone is watching.

These are normal operating conditions for agents that do background work.

This lesson is about building agents that survive those conditions: agents that checkpoint state, resume safely, manage context across long runs, report progress, enforce budgets, and handle work that outlives a single HTTP connection.