Agentic Loop Failure Modes in High-Acuity Settings — Preliminary Findings

Preliminary findings from a 6-month observational study of AI agent behavior in ICU and step-down unit environments. Key findings: (1) Loop abandonment — agents halt unexpectedly when encountering ambiguous sensor data, occurring in 3.2% of monitored sessions; (2) Confidence inflation — agents report high confidence scores on tasks outside their training distribution; (3) Escalation fatigue — nurses begin ignoring escalation alerts after repeated false positives, the most dangerous failure mode identified. Recommendations include mandatory confidence calibration audits, tiered alert systems with severity weighting, and regular nurse-AI calibration sessions. Full paper pending peer review.