Harness-S trains agent policies across a pool of sibling harnesses

A new reinforcement-learning recipe called Harness-S puts an agent policy through a pool of specialized 'sibling' execution harnesses during training, then deploys that same policy under the original base harness at inference time. The arxiv preprint reports gains across three agent benchmarks, AppWorld, SWE-Gym and DBBench, without publishing per-benchmark numbers in the abstract.

The method keeps a fixed 'Sibling Harness Pool' grown by what the authors call 'Directional Harness Evolution.' When all initial rollouts under the base harness fail, the pool is invoked, and trajectories collected under auxiliary harnesses replace the failed group for reinforcement learning. An 'Adaptive Harness Router' picks which Top-k variants to try using 'Harness Behavior Profiles and voting across multiple LLMs,' and a policy-aware re-anchoring step refreshes those profiles as the policy learns.

The abstract's bottom-line claim is that 'the trained policy substantially outperforms policy-only training when returned to the original base harness.' The word 'substantially' is the paper's own; no percentages appear in the abstract. The authors add that 'Reusing the router for Rank-1 inference provides additional gains at the same executor rollout count,' suggesting the routing mechanism carries over to serving time.

The abstract does not publish per-benchmark deltas, headline percentages, or which base model the recipe was trained on.