继续还是重启？编码智能体早期失败预测的局限
原标题：Continue or Restart? The Limits of Early Failure Prediction in Coding Agents
AI 摘要
该研究基于两个公开语料库的147,002条执行轨迹，覆盖四种模型-框架组合，评估编码智能体早期失败预测的可行性。研究发现，在5k-20k token的早期阶段，任务内区分度接近随机水平（AUC约0.50），早期前缀仅能小幅提升跨任务排序。在75k token预算下，学习到的重启策略相比随机重启最多提升1.98个百分点，但增益几乎全部来自任务间停止分配，任务内尝试级选择贡献接近零；无中断继续在所有四种设置下均以更低平均token成本取得更高解决率。
正文节选
Continue or Restart? The Limits of Early Failure Prediction in Coding Agents Abstract Early stopping promises to reduce the computational cost of autonomous coding agents by terminating unpromising runs before completion. Evaluating when to abandon an active run, however, requires separating the baseline difficulty of a software issue from the trajectory of an individual attempt. We study this distinction across 147,002 execution traces from two public corpora spanning four model-framework setti