Glossary
Model drift
Also called: model deprecation, prompt drift
Definition
Model drift is when an agent's output quality changes without its configuration changing, because the underlying model, an API, or a data source changed instead.
In practice
It is the failure mode most often misdiagnosed as a prompt problem. A model version is retired and traffic moves to its successor, an API renames a field, or a site the agent reads redesigns its markup. The agent is running perfectly and doing something subtly different. Drift is hard to catch because it degrades rather than breaks: the brief still arrives at 7am, it is just no longer worth reading. The only reliable defence is a habit rather than a tool, namely reading a sample of real output regularly, and re-testing deliberately whenever a model or dependency changes.
Read more
Nearby terms
All 18 termsRun agents without the homework
Qoren runs OpenClaw, Hermes, and Codex agents in managed environments, with the restarts, secrets, schedules, and spend caps handled for you.
Start free trial