How do you stop an LLM agent from drifting in production?
You do not prevent drift, you detect it early enough to act. Pin model versions explicitly, run a fixed canary eval set on every deploy and on a daily schedule, and alert on the change in score rather than its absolute value. Systems that drift silently are systems with no scheduled measurement.
Drift arrives from three directions and each needs a different sensor. Provider drift, where the model behind an unpinned endpoint changes underneath you, is caught by the daily canary run. Data drift, where the input distribution moves, is caught by monitoring embedding distributions of incoming requests against a reference window. Prompt drift, where a hundred small edits accumulate, is caught by the regression budget in CI.
Pin versions wherever the provider allows it, and treat a version bump as a deploy with its own eval run — not as a background event you find out about from users.
Keep the canary set frozen. The instant you edit it to make a red build go green, you have lost the baseline that made it useful.
Written by Binary AI Labs · Reviewed