StarHarness: Evolving Harnesses with Stratified Search for Enterprise Environments
Researchers presented StarHarness, a framework that evolves environment-specific agent harnesses while leaving underlying model weights unchanged. The harness can incorporate adjustments to prompts, task framing, tool interfaces, skills, providers, subagent structures, and agent-loop configurations. It builds a compact evolution pool by grouping tasks according to how baseline systems fail on them.
For robotics, drone, and automation operators, the method offers a way to adapt coordination logic across different sites or fleets without retraining core models. This targets enterprise settings where tasks vary by environment and where repeated failures on specific behaviors can be used to guide configuration changes.
The framework's reliance on stratified failure data for pool construction means deployment would require systematic logging of initial performance across representative tasks.