Large language model (LLM) agents acquire task-specific capabilities by loading reusable skill documents into a bounded context window. However, current methods for selecting skills independently without considering the impact on performance and token cost can result in redundant or poorly chosen skills wasting scarce tokens. Researchers Yu Chen et al. have developed an algorithm called Best Prefix Selection (BPS), which provides a polynomial-time solution to select a skill set under a hard token budget to maximize benefit while minimizing context penalty. The authors claim that BPS outperforms existing methods, reaching 73% task success on 28% fewer tokens. --- Why it matters: This matters because it addresses the issue of inefficient skill selection in LLM agents, which can lead to wasted resources and decreased performance. By providing a provable bicriteria guarantee, this research offers a more reliable approach to selecting skills for these models. Source: https://arxiv.org/abs/2608.19993