SkillApt loads only 31.5% of retrieved AI skills

SkillApt splits skill retrieval from skill activation, then decides whether to load each retrieved skill instead of assuming it belongs in context.

Most agent frameworks treat those as the same step. A retriever finds a SKILL.md file that looks relevant, and the agent loads it into context no matter what. SkillApt's tests show that habit is expensive. Loading only 31.5% of retrieved skills still hit 0.838 accuracy on SRA-Bench, identical to loading every top BM25 match, while cutting token use 74.3%.

For each retrieved skill, SkillApt runs matched executions with and without loading it. It then uses outcomes from similar past states to decide LOAD or ABSTAIN, per the SkillApt paper.

What the paper doesn't say is how the with/without comparison holds up outside SRA-Bench, or what running those matched executions costs in compute at scale. If the deciding mechanism itself needs meaningful inference budget, the net savings could shrink in production.

Pull the skills your own agent retrieves most often and count how many loads actually changed the output. If most of them sit unused, that's context budget recoverable without losing accuracy.