SkillApt Matches BM25 Top-1 Accuracy While Loading Retrieved Agent Skills Only 31.5% of the Time and Cutting Tokens 74.3%
arXiv·medium signal
arXiv 2609.26863 (22 Sep) separates skill retrieval from skill activation. It runs matched WITH/WITHOUT executions and uses outcomes from similar past states to decide LOAD or ABSTAIN per retrieved skill. On SRA-Bench it held 0.838 accuracy, equal to always loading the BM25 top hit, while cutting activation from 100% to 31.5% and mean tokens by 74.3%, which applies directly to anyone injecting SKILL.md files into agent context.