Dev.to
5/8/2026

My agent worked yesterday. Today it's possessed.
Short summary
LLM agents can drift silently—token usage spikes 50x, tool sequences change, outputs morph—without errors or crashes. Drift detection builds rolling baselines from normal sessions and fires alerts when behavior diverges statistically (Z-score > 2.0, Jaccard distance > 0.4), catching anomalies in seconds rather than weeks. tokenjam provides in-process drift detection with zero API keys, enabling you to monitor agent behavior the same way you measure deterministic systems.
- •Silent agent drift (behavior changes without crashes) is hard to diagnose but detectable via statistical baselines
- •Drift detection compares Z-scores and tool-call patterns against rolling history to alert on anomalies
- •tokenjam enables in-process monitoring with zero API keys; minimal configuration required
Generated with AI, which can make mistakes.
Is this a good recommendation for you?



