Agent failures are often blamed on the policy, but the execution harness can be part of the problem: different harness variants fix different failure modes, and a variant’s usefulness can change as the policy improves. Harness-S turns that variability into training signal, aiming to transfer behaviors prompted by auxiliary harnesses into one policy that can still work with the original setup.
Key Findings
- Keep complementary fixes, not one universal winner. Directional evolution builds a pool of sibling harnesses with distinct repairs, selected for complementary validation coverage. This treats harness diversity as a resource rather than searching for a single best configuration.
- Use auxiliary harnesses only after base attempts fail. The policy makes fresh attempts under selected siblings only when all its initial base-harness attempts fail; those trajectories replace the failed group for reinforcement learning. Successful base attempts remain the training group.
- Train for transfer, then optionally route at inference. On AppWorld, SWE-Gym, and DBBench, the trained policy outperforms policy-only training when evaluated under the original harness. Task-specific routing to a sibling can provide further gains without increasing the executor rollout count.
Method in Context
The sibling pool is created through successive rounds of harness editing based on recurring failure causes, then held fixed during policy training. A router selects auxiliary variants using harness behavior profiles and multi-LLM voting; policy-aware profile re-anchoring refreshes the selection evidence as learning changes the policy. The key distinction from simply choosing a better harness is that rescue trajectories are intended to teach a shared policy, not just improve execution for the duration of one task.
Who It Suits—and the Trade-off
This is most relevant to researchers training tool-using agents whose failures vary with prompts, tools, memory, or execution setup, and who want improvements to persist under a stable deployment harness. It is less compelling when a single harness already covers the task distribution, or when building and maintaining sibling variants and rescue rollouts is too costly. The reported results span three benchmarks, but the available summary does not provide numerical effect sizes, so the gains should be judged from the paper’s detailed tables.