What is drift in an AI agent?

Drift is any change in an agent’s behavior or output quality that happens without anyone deliberately changing what it’s supposed to do, whether the cause is a model provider updating a model version underneath you, a prompt someone tweaked, or the world the agent describes moving on without it. It’s borrowed from two older machine learning terms, data drift, where the inputs a model sees start to look different, and concept drift, where the correct answer for the same input changes, but an agent can drift for reasons neither term was built for, like a tool’s API changing shape.

The practical difference from a regression is timing. A regression traces to one change you can point at; drift is what you call the same kind of quality drop when nothing in your own history explains it, which usually means the cause is upstream, in the model or the world, not in your code.