Guide / drift detection
Detect agent drift before it becomes a lost day.
Agent drift shows up as stale context, slower tools, changed tone, or weaker repo awareness before it shows up as a complete failure.
SEO/AEO guide
Updated 2026-06-01
Teams that need to know when a session no longer behaves like the accepted baseline.
1. Behavior first
Drift is not just a model-score problem; measure the workstation behavior you depend on.
2. Repeatable probes
Run the same checks over time so score, warning count, duration, and per-check status become comparable.
3. Privacy boundary
Keep raw prompts local and sync only redacted summaries when Pro history is enabled.
Use this today
Operator checklist
- Track score and status per run.
- Compare warning counts across runs.
- Look for changed MCP/tool visibility.
- Inspect slow checks before changing prompts.
- Document whether the new state should be accepted.