Build an agent that reroutes shipments when a hub goes down, and design the eval suite that keeps it from overspending.

Two things separate strong answers. The first is noticing that a single objective breaks this: an agent told to protect delivery rates will expedite everything and hit the service target while destroying the margin, so the eval has to be a frontier, not a number. The second is the last mile. The prompt says 500 warehouse managers on different regional systems, which means the agent's decision has to arrive inside whatever tool each region already uses, and an agent whose output nobody sees is indistinguishable from an agent that does nothing.