Agent skills are often shortened, by compression, rewriting with a language model or partial loading, to cut the cost of running LLM agents. We tested whether this works on 40 SkillsBench tasks. The agent received a curated skill either as written or in one of several shortened versions, each version ran on a small and a large model three times per task, and every run was priced at the provider's published rates. The shortened skills were up to 54% shorter, and cost per task stayed at or slightly above that of the original skill, because the skill is a small and mostly cached part of the text the agent processes. Shortening by a program that keeps the original sentences kept the success rate at the level of the original skill. Rewriting by a language model lowered success on the small model, from 42.5% with the original skill to between 29.2% and 37.5%, and a stronger rewriting model gave the same result. The largest change came from the model that runs the agent: moving to the large model raised success by 18 to 27 points at 4.6 to 6.5 times the cost. The results favour using curated skills as written and spending effort on the choice of agent model.

翻译：暂无翻译