SkillOS: RL-Trained Skill Curator Enables Self-Evolving LLM Agents

arXiv·medium signal

SkillOS trains a skill curator via composite rewards and task-stream grouping so LLM agents learn which skills to keep, refine, or discard from experience. Skills evolve into structured Markdown files encoding meta-skills. The learned curator generalizes across different executor architectures and task domains, outperforming both memory-free and heuristic-based skill management baselines.