SkillSV Values Agent Skills for Optimization and Pruning

As AI agents become more sophisticated, their "skills" are often composed of numerous structured internal units like rules, examples, or scripts. Understanding the individual contribution and value of these units is crucial for optimization, but traditional valuation methods fall short due to the inherent structure and dependencies within skills. This research introduces SkillSV, a novel framework based on Shapley values, specifically designed for skill valuation. SkillSV compiles an agent's skill into its constituent units, their dependencies, and hierarchical relationships, ensuring that only valid counterfactual skills are evaluated. It employs paired deletion and length-neutral padding to isolate the content value from context costs, estimating these values through a rollout-budgeted estimator. The framework demonstrates its effectiveness across various agentic benchmarks by accurately recovering unit interactions, preserving aggregate skill lift, and providing guidance for safe pruning and compression of agent skills.