What Should a Skill Remember? Quality--Cost Trade-offs in Cost-Aware Skill Rewriting for Language Model Agents

The paper presents a framework for cost-aware skill rewriting in language model agents, highlighting the quality-cost trade-offs associated with different skill structures. Using the SkillsBench benchmark, it demonstrates that applying strategies such as API/code anchoring and rule/formula anchoring can achieve an average reduction in total cost by 7.0% and downstream agent-token cost by 6.0%, while maintaining verifier quality. This research emphasizes the importance of skill design as a critical component of operational knowledge engineering, rather than merely prompt compression, which is crucial for practitioners aiming to optimize resource efficiency in LLM applications.