Output-distribution drift
Also known as: model drift, agent drift, output drift
A measurable shift in the distribution of an AI agent's outputs across a defined evaluation set, sampled at minimum weekly. Drift can manifest as a change in classification accuracy, a change in tone or content distribution, a change in tool-use patterns, or a change in refusal behaviour — without any change in the deployment configuration. One of the four MTTD-for-Agents tripwires the publication recommends as a 2026 enterprise procurement bar.
Output-distribution drift is the slowest of the four MTTD tripwires to fire and the most under-instrumented in 2026 enterprise deployments. Most programmes treat the eval set as a one-time procurement artifact rather than a standing measurement. The drift detection pattern: define an evaluation set at procurement, run it weekly under production conditions, compute distribution-shift statistics on the agent's outputs, alert on sustained shift. The vendors that support customer-supplied weekly evaluation sets natively are at the 2026 enterprise bar; the ones that don't are not.