harness-optimization

Retrospective Harness Optimization (RHO) is a self-supervised method that improves LLM agent performance using only past trajectories, achieving a 78% pass rate on SWE-Bench Pro without external grading.