SHarP: Saliency-based Pruning of Agent Harnesses

SHarP optimizes agentic workflows by pruning redundant tools, instructions, and modules based on their saliency relative to task performance. The method reduces token overhead by identifying components that contribute little to execution success, enabling a more efficient balance between system complexity and operational costs in complex LLM harnesses.