Agent drift or model drift is a measurable change in behavior over time. Model drift can come from a model update or distribution shift. Agent drift can also come from prompt changes, tool changes, retrieval changes, memory changes, policy changes, or user behavior shifts.
For production agents, drift should be tracked with online evals, slice-based metrics, and regression datasets. The question is not only "did the model change?" It is "did system behavior change in a way users notice?"