Harness-R1 trains a 9B model to patch the agent harness itself from failure trajectories, adding 9.3 points without touching the agent
Posted 2026-08-03 (arXiv 2608.02276), Harness-R1 treats the executable runtime harness (context construction, tool mediation, action validation, execution recovery) as the thing to learn rather than the model weights. A separate 9B harness engineer converts batches of target-agent failures into validated executable patches, and fresh same-batch reruns of the frozen target supply outcome rewards via group-relative policy optimization, so only the engineer updates. Across WebShop, ALFWorld, and DBBench it lifts vanilla Qwen3.5-9B from 44.3% to 53.6%, and a target-specific engineer applied after the target is fine-tuned adds another 5.0 points (59.2% to 64.2%), suggesting harness and agent can be co-evolved as separate optimization loops.
Source
↳ Follow the thread