Anomaly Patterns

Threshold alerts catch sudden failures. Anomaly detection catches slow degradation — the kind where every individual data point is within normal range, but the trend is moving in the wrong direction.

Five anomaly patterns to monitor. The Slow Leak: a metric drifting 0.5% per day — invisible day-to-day, but 15% degradation over a month. Detected by linear regression on 14-day rolling windows. The Phase Shift: a metric that suddenly establishes a new baseline without an obvious cause. Average cycle time was 47 minutes for three weeks; now it is 53 minutes. No single data point triggered an alert, but the system has changed. The Oscillation: a metric that swings between two states — high quality on Monday/Tuesday, degraded quality Thursday/Friday. Indicates a systemic pattern, not random variance. The Correlation Break: two metrics that normally move together (throughput and quality) suddenly diverge. Throughput stays constant while quality drops — the agent is producing faster but worse. The Cascade: a degradation in Agent A's metrics followed 24-48 hours later by degradation in Agent B and C's metrics. The root cause is upstream; the symptoms are downstream.