Researchers have published SkillsBench, the first standardized benchmark for measuring whether "Agent Skills" — structured knowledge packages bolted onto LLM agents at runtime — actually improve performance.
The benchmark covers 87 tasks across 8 domains and tests each one under matched conditions: with curated Skills and without. Across 18 model-and-harness combinations, adding Skills pushed the average pass rate from 33.9% to 50.5%, a gain of 16.6 percentage points. Two findings cut through the noise: focused packages of three modules or fewer outperform larger, exhaustive bundles, and smaller models equipped with Skills can match larger models running without them.
Agent Skills are everywhere now — every major agentic framework ships some version of pluggable procedural knowledge — but until this paper there was no agreed-upon way to verify whether they delivered. SkillsBench introduces paired evaluation as the baseline methodology: run the same task twice, with and without Skills, and measure the delta. That makes claims about Skill efficacy falsifiable, which matters more than it sounds in a space where "augmented performance" tends to mean a vendor's benchmark on a vendor's task set. The smaller-model finding also has cost implications: if a focused skill set closes the gap between a cheap model and an expensive one, the build-versus-scale tradeoff looks different.
Of course, 87 tasks is a narrow slice of what agents encounter in production, and deterministic verifiers are easiest to construct when tasks have clean right answers. Whether the industry adopts SkillsBench as a shared standard or files it under academic citation fodder is the real test.