临床症状检测中自主代理工作流程的优化不稳定性

Autonomous agentic workflows that iteratively refine their behavior show great promise, yet their failure modes are not well characterized. This research investigates optimization instability, a phenomenon where continued autonomous improvement paradoxically degrades classifier performance. Using Pythia, an open-source framework for automated prompt optimization, three clinical symptoms with varying prevalence were evaluated: shortness of breath (23%), chest pain (12%), and Long COVID brain fog (3%). Validation sensitivity was observed to oscillate between 1.0 and 0.0 across iterations, with severity inversely proportional to class prevalence. At a prevalence of 3%, the system achieved 95% accuracy while detecting zero positive cases, highlighting the potential risks inherent in the optimization process.