SkillOptimizer: Agent Skill Optimization Through Subskills Without Task Supervision

Nicholas Crispino, Shubham Gaur, Xuefang Yang, Angela Yu, Berat Ercevik, Clara Sapugay, Yujin Potter, Dawn Song, Chenguang Wang

Keywords: LLM agents, agent skills, agent harnesses

TL;DR: SkillOptimizer improves agent skills without using downstream tasks — it decomposes each skill into a DAG of subskill decisions, then iteratively refines via synthetic coverage-driven prompts.

Abstract: Agent skills have become a primary mechanism for extending LLM agent capabilities by providing raw instructions, resources, and domain knowledge. Yet skill quality in the wild is uneven: skills do not always deliver on all the workflows promised by their descriptions, and existing methods for improving skills automatically require downstream task data and verifiers that are often unavailable in practice. We present SkillOptimizer, a pipeline that improves skills without consulting any downstream task or verifier. SkillOptimizer decomposes each skill into a directed acyclic graph of subskills, synthesizes coverage-driven prompts from the DAG structure, and runs iterative optimization cycles to improve each subskill. Applied to 192 skills spanning 83 tasks, SkillOptimizer improves task pass rate over raw baselines across three agent families, establishing that current skills leave coverage value that structured synthetic optimization can recover.

Track: Regular Paper (9 pages)

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 271