SkillOS: Learning Skill Curation for Self-Evolving Agents

Introducing SkillOS, a technology curation framework for self-evolving AI agents. If existing agents are just a one-off solution that doesn't take advantage of past experience, SkillOS learns how to build and manage a practical **SkillRepo** through reinforcement learning. The system consists of a fixed execution agent and a learnable curator agent, which inserts, modifies, deletes, and optimizes the repository based on the results of the work. Research has shown that SkillOS has significantly improved operational efficiency by increasing the success rate of agents' tasks while reducing unnecessary exploration. In particular, the learned curator has demonstrated excellent universality and generalization ability in various working environments and model engines. As a result, the paper presents a new methodology that continuously develops the intelligence of agents by distillating experience into structured techniques. https://arxiv.org/pdf/2605.06614