Data From 100+ AI Agents Shows They Fail by Silent Drift, Not Errors
AI Summary: The authors analyzed alert data from over 100 production AI agents across dozens of companies during a seven‑month period, grouping alerts into output‑quality, token‑spend/latency, runaway‑loop, and trajectory/validation checks. They found that 89 % of agents triggered at least one alert, with a median time to the first alert of six days and half of the alerts occurring within a week of monitoring start. Output‑quality, token‑spend, and latency alerts fired early and frequently (≈75‑83 % of agents) while error alerts were rare and typically the last to fire. Alerts for token spikes almost always coincided with drops in output quality, and runaway‑loop issues tended to become chronic, highlighting a monitoring focus on performance metrics and a gap in governance and safety oversight.