AI agents rely on a harness of skills, tools, and workflows to solve complex problems, and continually improving this harness is essential for adapting to new tasks. Existing optimization methods typically require ground-truth validation sets, yet such labeled data is difficult to acquire in practical deployment settings.
We introduce Retrospective Harness Optimization (RHO), a self-supervised method that optimizes the agent harness using only past trajectories. RHO selects a diverse coreset of challenging tasks from past trajectories and re-solves them in parallel. The agent analyzes these rollouts using self-validation and self-consistency, then generates candidate harness updates and selects the most effective one by its own pairwise self-preference.
We evaluate RHO across three diverse domains spanning software engineering, technical work, and knowledge work. Notably, a single optimization round improves the pass rate on SWE-Bench Pro from 59% to 78% without any external grading. Our analysis shows that RHO effectively targets prior failure modes: the optimized harness alters the agent's behavior patterns and sustains higher accuracy during long-horizon sessions.
Most harness optimizers iterate against a labeled validation set — but real deployments rarely have one. A deployed agent does, however, produce a continuous stream of unlabeled trajectories. RHO turns those into harness improvements in three label-free stages:
A determinantal point process (DPP) picks a small, difficulty-diverse subset of past tasks to re-solve.
Each coreset task is re-solved G times in parallel, yielding two diagnostic signals: self-validation within a trajectory and self-consistency across trajectories.
The agent samples N candidate harness edits and keeps the one its own pairwise self-preference ranks highest over the baseline.
Validation-feedback methods repeatedly score harness edits against labeled data. RHO instead reflects on past trajectories in a single retrospective pass — replacing the external grader with the agent's own self-validation, self-consistency, and self-preference.
Held-out pass rate after a single optimization round (Codex + GPT-5.5), against feedback-free baselines under a matched agent-call budget:
| Method | Harness surface | SWE-Bench Pro | Terminal-Bench 2 | GAIA-2 |
|---|---|---|---|---|
| Vanilla Codex | — | 0.59 | 0.71 | 0.29 |
| Dynamic Cheatsheet | Skills | 0.62 +0.03 | 0.73 +0.02 | 0.30 +0.01 |
| ReasoningBank | Memory | 0.61 +0.02 | 0.73 +0.02 | 0.28 −0.01 |
| Sleep-time Compute | Memory | 0.64 +0.05 | 0.73 +0.02 | 0.32 +0.03 |
| RHO (ours) | Skills + Tools | 0.78 +0.19 | 0.76 +0.05 | 0.37 +0.08 |
RHO also beats Meta-Harness, a validation-feedback optimizer, at a matched single-round budget (0.78 vs 0.62 on SWE-Bench Pro) — without ever touching ground-truth labels.
@article{pan2026rho,
title = {Retrospective Harness Optimization: Improving LLM Agents
via Self-Preference over Trajectory Rollouts},
author = {Pan, Wenbo and Liu, Shujie and Lin, Chin-Yew and Zeng, Jingying
and Tang, Xianfeng and Zhou, Xiangyang and Lu, Yan and Jia, Xiaohua},
journal = {arXiv preprint arXiv:2606.05922},
year = {2026}
}