Definition
Plain language
A way for an AI agent to improve itself by studying its own past attempts — with no answer key — and rewriting its own tools.
As stated in the literature
A label-free method that selects hard, diverse past trajectories, re-solves and diagnoses them via self-validation and self-consistency, proposes harness code edits, and selects among candidates by self-preference (pairwise comparison against the baseline) under a strict-improvement gate.
Also called: RHO
Why it matters: It lets an agent improve itself from its own history even when no labeled correct answers are available.
For example, an agent reviews its past failed attempts at a task, re-solves them and judges which version went better, and rewrites its own tools accordingly — all without an answer key.
Heard on the show
“The paper is called "Retrospective Harness Optimization: Improving LLM Agents via Self-Preference over Trajectory Rollouts.”Episode 120 — How an AI Agent Rewrites Its Own Tools, Without an Answer Key