The paper introduces Retrospective Harness Optimization (RHO), a self-supervised method that optimizes an AI agent’s harness (its skills, tools, and workflows) using only its own past trajectories. RHO selects a diverse coreset of challenging tasks from those trajectories, re-solves them in parallel, and analyzes the rollouts using self-validation and self-consistency checks.
In a single optimization round on SWE-Bench Pro, RHO improved the pass rate from 59% to 78% without any external grading. The method effectively targeted prior failure modes and sustained higher accuracy during long-horizon sessions. Evaluations across software engineering, technical work, and knowledge work showed consistent benefits.
This approach is interesting because it enables continuous improvement of AI agents in deployment settings where labeled validation data is scarce or absent, reducing the reliance on human annotation.