Search for a command to run...
Retrospective Harness Optimization: Improving LLM Agents via Self-Preference over Trajectory Rollouts