Agent reinforcement learning increasingly uses full execution harnesses, with a focus on how multiple harnesses influence policy optimization. The study isolates two choices: exposing policies to multiple harnesses and comparing rewards within a relative-advantage group.
Using a Qwen3-8B supervised warm start, the evaluation replayed frozen task-harness records across different systems, applying two rules for group-relative policy optimization: Within and Cross. The evaluation involved 24,000 sealed tests, revealing that the evaluation harness itself has a major impact on solve rates, increasing the mean from 2.14% to 9.27%. The grouping rule (Within vs. Cross) had a smaller, less consistent effect.
Results indicate that the largest gains occur on the same source harness, and the advantage from Cross-harness credit can be identified by an out-of-fold classifier. However, this advantage does not translate into increased portability of capabilities across different harnesses, and re-collecting data on-policy does not alter this outcome. The report emphasizes that multi-harness RL results should specify the grouping boundary and test on unseen harnesses to assess transferability.
Source: https://arxiv.org/abs/2609.04518