New Algorithm Optimizes LLM Agent Skill Selection.

Large Language Model (LLM) agents rely on loading reusable "skill documents" into their limited context windows to acquire task-specific capabilities. However, current skill selection methods, often based on semantic relevance or greedy packing, lack quality guarantees and cost awareness, leading to wasted tokens and potentially degraded performance. This research models skill selection as an optimization problem: choosing a skill set to maximize a monotone submodular benefit while adhering to a token budget. The paper introduces Best Prefix Selection (BPS), a polynomial-time algorithm that offers the first known performance guarantee for skill selection—a bicriteria (1-1/e, 1) approximation. In experiments on a BigCodeBench variant, BPS achieved a 0.73 task success rate, significantly outperforming released skill routers, text retrievers, and the executor's own selection (0.20-0.52), while using 28% fewer tokens than the strongest router. This demonstrates a more efficient and effective way for LLM agents to acquire and utilize skills.