Time Horizons
Drift is computed on every observation run across three windows:| Window | Comparison period |
|---|---|
| Daily | Last 24 hours vs. baseline |
| Weekly | Last 7 days vs. baseline |
| Monthly | Last 30 days vs. baseline |
Drift Categories
Behaviour Drift
Behaviour Drift
Computed per intent, compared to baseline:
| Metric | What it measures | Scoring method |
|---|---|---|
| Cost | Average cost per trace | Z-score |
| Latency | Average trace latency | Z-score |
| Step count | Average number of steps an agent takes to complete a task | Z-score |
| Error rate | Share of traces with errors | Percentage change |
| Tool distribution | Which tools the agent calls and how often | Jensen-Shannon divergence |
Output Drift
Output Drift
Tracks average output length per intent. Surfaces when responses become noticeably shorter or longer than the baseline - a common signal of prompt regression or model behavior change.
| Metric | Scoring method |
|---|---|
| Output length | Percentage change |
Input Drift
Input Drift
Emitted alongside new-intent discovery. When a cluster of user inputs appears that does not match any known intent, Insights flags it as an input drift signal - indicating that the nature of requests hitting your AI has shifted.
Drift Severity
Each metric is scored automatically using the method appropriate for that signal type. You do not need to configure thresholds; Insights applies calibrated defaults and assigns a severity to each observation.Learn More
Intents
Understand how Insights classifies user requests by intent
Alert Rules
Set up alerts on top of the metrics Insights tracks