Researchers propose SkillCorpus to consolidate and evaluate open agent skills for LLM workflows

Agent skills—packaged as SKILL.md files that encode reusable procedural knowledge for LLM agents—are widely shared in public repositories but are fragmented, redundant, and uneven in quality. SkillCorpus addresses this by aggregating, curating, matching, and evaluating the open skill ecosystem at scale.

The framework filters approximately 821,000 crawled skills through a multi-stage pipeline into 96,401 skills organized by a 16-class taxonomy and three quality facets: utility, robustness, and safety. It pairs this corpus with a fine-tuned retrieval-and-selection stack designed to match task-relevant skills to agent workflows.

Evaluations across three benchmarks—SkillsBench, GDPVal, and QwenClawBench—two harnesses, and two open backbones include a frontier robustness check. Integrating SkillCorpus yields consistent performance gains across all three benchmarks, with the largest improvement observed on SkillsBench at +7.5 percentage points.

An operational analysis attributes the gains to two boundaries: a coverage boundary where the corpus supplies relevant skills, and a harness boundary where the retrieval and selection stack effectively surfaces those skills. The authors state that SkillCorpus is the first end-to-end account of when a curated, retrieval-served community corpus improves real agent tasks, and where it does not.

The dataset, models, and code are slated for release upon acceptance.