Shanghai Qijing's ADE-PRF predicts AI agent failures 8 hours ahead

Shanghai Qijing Digital Technology Co. has introduced the ADE Predictive Reliability Framework (ADE-PRF), a system designed to foresee reliability issues in long-horizon LLM multi-agent systems—an area where traditional infrastructure monitoring often fails. The framework aggregates 20 heterogeneous signals across five architectural layers into a single Trust Margin (TM) metric with a 39.2-point dynamic range. By running three parallel prediction methods (Exponential, Kalman, and a third unstated), ADE-PRF generates 8-hour forecasts. The Exponential method leads with a Mean Absolute Error of 1.228 and a Direction Accuracy of 76.8%, while 99.65% of predictions fall within a ±10-point tolerance.

Validation was extensive: the team conducted 380,227 predictions and 280,579 validations across six distinct agent profiles over 15 continuous days, plus seven sandbox-controlled experiments. A standout finding is the detection of “false prosperity”—a state where degradation is concealed by normal surface-level metrics, luring operators into a false sense of security. Upon integrating the ADE plugin, the Trust Margin immediately couples with ground-truth system states, with 16 of the 20 signal factors relying on data collected by the plugin. Comparative tests show the Exponential method consistently outperforms the Kalman filter. ADE-PRF offers among the earliest forward-looking reliability quantification for production LLM agents, giving teams proactive warnings instead of reactive post-mortems.