EvoSkills: Self-Evolving Agent Skills via Co-Evolutionary Verification

Source: arXiv Authors: Hanrong Zhang, Shicheng Fan, Henry Peng Zou, Yankai Chen, Zhenting Wang, Jiayu Zhou, Chengze Li, Wei-Chieh Huang, Yifei Yao, Kening Zheng, Xue Liu, Xiaoxiao Li, Philip S. Yu Date: 2026-04-02 Primary category: cs.AI All categories: cs.AI

Abstract

EvoSkills treats skills as structured multi-file artifacts and studies how agents can autonomously generate them. A Skill Generator iteratively refines candidate skills while a co-evolving Surrogate Verifier provides actionable feedback without access to hidden ground-truth tests. The architectural lesson is extremely direct: self-evolving workflows need a verifier lane, not just a mutation lane.