Harness-S: Learning from Sibling Harnesses for Generalizable Agent Training

Researchers introduced Harness-S, a method that uses sibling harnesses to selectively rescue failures of a base execution system and transfer capabilities into a single policy. The approach involves Directional Harness Evolution to create specialized sibling harnesses, Complementary Harness Rescue Training to invoke them when initial attempts fail, and an Adaptive Harness Router to select top variants using behavior profiles and voting across multiple LLMs.