A team ships a new agentic feature that runs without issue in the demo environment, but a week later in production, a routine cluster deploy causes the agent to instantly forget what it was doing, drop a half-completed task, and leave behind orphaned changes to systems of records and frustrated customers. Teams often blame the LLM first, but logs show the model often did exactly what it was supposed to do, requiring analysis of the execution layer underneath. The resilience of an agentic system depends on guarantees for its inputs and execution order, evaluated across four axes: Execution Durability (state lives in memory at the primitive end, where a crashed process destroys context; at the mature end, every step persists); Work Duration (primitive systems handle seconds within a single session; mature systems support durable timers, waiting, polling, periodic jobs, human approvals, and resumable work for days or months); Hosting and Isolation (dangerous operations run in provisioned, isolated environments with lifecycle management); and Quality of Service (mature systems handle backpressure and priority to avoid slowdowns and outages during concurrency spikes). Enforcing these guarantees requires end-to-end operational automation without relying on manual repair.