Skills Bench
person concept tool org talk claim — click a node to jump to its page; hover an arrow for the relation
Skills Bench is an evals|evaluation benchmark that indexes over 50,000 skills, used to measure the performance impact of skills on AI agents.
Overview
Philipp Schmid (Google DeepMind) describes Skills Bench as "a very popular and nice eval or benchmark" covering more than 50,000 skills. He uses it as a key reference point in the context of his argument that skills should not be shipped without proper evaluation.
Key Findings
According to Schmid, results from Skills Bench indicate that skills on average improve agent performance by roughly 15%. This figure serves as empirical grounding for the value of incorporating skills into agent pipelines, while also implying that unevaluated skills carry meaningful risk of regression or underperformance relative to that benchmark.