Retrospective Harness Optimization: Improving LLM Agents via Self-Preference over Trajectory Rollouts

2026-06-04

AI Summary

Current large language model agents rely on human-annotated validation sets for skill refinement, which hinders continuous iteration. This work proposes a self-supervised trajectory replay method that replays diverse, challenging tasks and leverages self-verification and self-consistency analysis to generate candidate strategies. A trajectory-level self-preference mechanism is introduced to select the optimal policy update without requiring external annotations, enabling autonomous optimization of skill composition. Evaluated on SWE-Bench Pro, the approach improves pass rates from 59% to 78% in a single round of optimization, effectively rectifying historical failure modes and maintaining higher accuracy on long-horizon tasks. To our knowledge, this is the first application of self-preference mechanisms to trajectory-level agent behavior improvement.

Abstract

AI agents rely on a harness of skills, tools, and workflows to solve complex problems. Continually improving this harness is essential for adapting to new tasks. However, existing optimization methods typically require ground-truth validation sets, yet such labeled data is difficult to acquire in practical deployment settings. To address this problem, we introduce Retrospective Harness Optimization (RHO), a self-supervised method that optimizes the agent harness using only past trajectories. Specifically, 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. Furthermore, our analysis demonstrates that RHO effectively targets prior failure modes. As a result, the optimized harness alters the agent's behavior patterns and sustains higher accuracy during long-horizon sessions.

Problem

• harness optimization
• LLM agents
• self-supervised learning
• trajectory rollouts
• validation-free optimization

Innovation

• Retrospective Harness Optimization
• self-supervised learning
• trajectory rollouts
• self-preference
• LLM agents