Harness-S: Training Smarter AI Agents Using Sibling Execution Systems

Background Training AI agents to handle complex coding and tool-use tasks usually relies on a fixed execution harness. However, every tweak to these scaffolds tends to fix some errors while breaking others, making it hard to find a universally superior setup. Details To solve this, researchers developed Harness-S. It builds a pool of specialized sibling harnesses along different repair directions. When the primary base harness fails completely on a task, the system routes the problem to these auxiliary harnesses to generate successful trajectories. An Adaptive Harness Router uses behavior profiles and LLM voting to select the best helper harness. The agent then learns from these rescued runs, effectively absorbing the helper behaviors into its own core policy. Impact Experiments across benchmarks like AppWorld, SWE-Gym, and DBBench show that the resulting policy greatly outperforms standard reinforcement learning methods when returning to the original base system. This proves that diverse execution environments can successfully transfer useful skills directly into the AI model itself.