An anomaly detector learns or defines a baseline for normal behavior, then scores new events according to how unusual they are. In artificial intelligence workflows, it can monitor spending, tool use, output rates, errors or other operational signals.
An unusual event is not automatically harmful, and a harmful event may resemble normal activity. Thresholds, context and human review are needed to balance false alarms against missed incidents and to update the baseline as legitimate behavior changes.
Acronyms and aliases
outlier detection synonym
General terms
Related terms
Frequently asked questions
What can anomaly detection monitor in an artificial intelligence system?
It can monitor cost, token use, tool activity, error rates, loop length, data access and other operational signals.
Does an anomaly prove that something is wrong?
No. It is a signal that behavior differs from the expected pattern and should be evaluated with context.
