[2026-06-04] From Failed Trajectories to Reliable LLM Agents: Diagnosing and Repairing Harness Flaws
HarnessFix 先把 raw trace 编译成带数据流、控制流和代码锚点的 HTIR，再做归因、生成 scoped patch 与回归验证。本文复盘真实 AppWorld case、四 benchmark…
# Agent Learning
[2026-06-08] Self-Harness: Harnesses That Improve Themselves
Self-Harness 让同一个固定模型既执行任务，也从失败中提出并验证自己的 harness 修改。本文拆解 weakness mining、最小编辑、双 split promotion gate、九组结果，以…
[2026-08-03] Harness-R1: Learning to Edit Executable Runtime Harnesses from Agent Failure Trajectories
Harness-R1 把 harness 修补训练成一个 9B 模型的专门能力：从 failure packet 生成 lifecycle-wide patch，再用冻结 target agent 的真实重跑收益做…
