NVIDIA puts agent skills through live tests

NVIDIA’s open-source SkillEvaluator found that verified agent skills improved average specialized-task scores by 31 points across its first benchmark of more than 300 skills for 30-plus products, according to NVIDIA’s evaluation report. The paired live runs measure the same task with and without a skill; a companion ACES study found static structural and LLM-judge scans correlated at only ρ = 0.14 across 145 skills, while live evaluation captured discovery, workflow, and tool-use effects that documents cannot see. Teams building internal skill libraries should add executable paired tasks before trusting lint or rubric scores, extending the archive’s evidence that skill libraries degrade after 20 entries.