Presentation at the 1st Agent Skills Workshop at ACM CAIS 2026: Scalable and Efficient Agent Skill Retrieval System

Fangzhou Li presented the poster for SkillFlow: Scalable and Efficient Agent Skill Retrieval System, co-authored with Pagkratios Tagkopoulos and Prof. Ilias Tagkopoulos, at the 1st Agent Skills Workshop at ACM CAIS 2026 on 5/26/2026.

About this work: AI agents get sharper with the right skill and worse with the wrong ones, so SkillFlow sets out to find the few that matter out of tens of thousands. As the first open, multi-staged retrieval system for SKILL.md bundles, it runs a four-stage funnel that narrows ~36,000 community-contributed GitHub skills down to a handful in tens of seconds, trading breadth for precision at each step. Beyond the pipeline itself, SkillFlow contributes an open, reusable skill library and a thorough analysis: where the right skills exist, it lifts agent success by 78%, and where they fall short, gains disappear, pinpointing the skill library’s coverage and quality as the real ceiling on what agents can do. The findings translate directly into actionable guidance for improving both how agents search for skills and how those skills are written.

Reference: Li, Fangzhou, Pagkratios Tagkopoulos, and Ilias Tagkopoulos. “SkillFlow: Scalable and Efficient Agent Skill Retrieval System.” First Workshop on Agent Skills (2026). doi: 10.48550/arXiv.2504.06188

GitHub: https://github.com/IBPA/skill-flow