The classic machine-learning failure modes where a model's performance decays as the world around it changes.
Data drift is when the input data shifts from what the model saw in training. Concept drift is when the relationship between inputs and the right answer changes. Model drift is the resulting decay in accuracy over time. These are all about a statistical model degrading. AI agent drift is broader: an agent can drift from a prompt change, a new tool, or a policy update, not just a shift in data, and it is measured against what the business intended, not only against a benchmark.
Related: AI Agent Drift, Agent Behavioral Baseline
See also: What Is AI Agent Drift?