A 9B model just outperformed a 397B one at repairing agents， and the reason is not size.

The 9B model had been specifically trained with supervised fine-tuning and online RL， receiving feedback from whether its executable patches actually improved the frozen target agent.

Harness-R1 trains one agent to rewrite the execution layer around another frozen agent.

Because a production agent's failed runs may already contain the data needed to improve its runtime.

Harness-R1 uses those failures to train a separate 9B harness engineer that edits the executable layer around a frozen target agent.

From each failure batch， it writes validated code hooks for episode setup， pre-decision guidance， pre-action mediation， or post-feedback recovery.

The patched target reruns the same tasks， and the measured change in task reward updates only the engineer through online GRPO.

So the target model never drifts under the reward signal， while the editor learns which patches improve execution rather than which ones merely look reasonable.

• arxiv. org/abs/2608.02276

Title： "Harness-R1： Learning to Edit Executable Runtime Harnesses from Agent Failure Trajectories"