Optimizing Agent Skill Loading Reduces Token Costs Significantly.

The way AI agents load their skills can significantly impact token costs, especially in large language model (LLM) applications. This study investigates four distinct skill loading methods: "Full" (pre-loading everything), "Skill Block" (loading relevant blocks), "Reference" (loading references), and "Hybrid" (combining approaches). The goal was to determine which method optimizes token usage without sacrificing performance across various tasks, including SearchQA, SpreadsheetBench, ALFWorld, ScienceWorld, and SynthProc. The findings indicate that there is no single universally superior method, but conditional loading strategies, particularly "Skill Block" and "Hybrid," offer substantial reductions in input token usage. For instance, Hybrid reduced input by 27.4% on SearchQA and 39.8% on SpreadsheetBench. On more complex, multi-turn tasks like ScienceWorld and SynthProc, Skill Block and Hybrid achieved even greater reductions, up to 73.0% and 66.6% respectively. Crucially, paired outcome tests showed no detectable quality differences, suggesting these efficiency gains do not come at the expense of performance. The research concludes that conditional loading is most beneficial when only portions of a skill are needed per turn.