Harness-R1: A 9B Model Outperforms 397B in Repairing Agents via Online RL
rohanpaul_ai · x · 2026-08-05
A new paper, 'Harness-R1,' introduces a novel approach where a specifically trained 9B model successfully improves a massive 397B target agent.
Core Mechanism:
- Harness-R1 trains a separate 'engineer agent' dedicated to rewriting and editing the executable runtime harness around a frozen target agent.
- It extracts data from the target agent's failure trajectories, writing validated code hooks for episode setup, pre-decision guidance, or post-feedback recovery.
- Using online GRPO, the 9B editor model is updated solely based on the actual measured change in task reward.
This method ensures the large target model never drifts under the reward signal, while the smaller editor learns which patches genuinely improve execution rather than just appearing plausible.
Related event: Harness-R1 Enables Agents to Self-Repair from Failure Trajectories(3 posts)→
More from coding & agent
- Agent memory design question: what deserves a place in permanent memory? — Miserable-Shock3552 · 2026-09-22
- Dev demos near real-time editing with Gemini Flash Lite, plans open-source Chrome extension — Saboo_Shubham_ · 2026-09-22
- Local MLX vs hosted API: two AI agents battle in Chrome's T-Rex game — usamawahabkhan · 2026-09-22
- A silent watcher race dropped agent logs for a month: full post-mortem of a tailing bug — ClaudeCdGuy · 2026-09-22
- D-RAC cuts enterprise document chunking tokens by 95.7%, cost by up to 85.6% — Yellow-AI-NLP · 2026-09-22
- Did the Action Happen? Designing Confirmation Logic for Agent Tool Calls That Time Out — felix_baron · 2026-09-22