SkillOS: Learning Skill Curation for Self-Evolving Agents (arxiv.org)

A new approach called SkillOS was announced, designed to enhance self-evolving AI agents by enabling them to learn and curate skills from previous interactions. Traditional systems often rely on manual skill curation or limited training that fails to adapt over time. SkillOS proposes a dual-system architecture where a frozen agent executor retrieves and applies learned skills, while a trainable skill curator updates an external SkillRepo based on gathered experiences. This setup allows for long-term learning, improving the efficiency and effectiveness of AI agents across various tasks. The significance of SkillOS lies in its ability to create a continuous learning ecosystem for AI agents, addressing the limitations of current methods that struggle with complex, long-term skill curation. By utilizing composite rewards tied to task dependencies, the SkillOS framework enables more targeted skill application and fosters the evolution of structured skills stored in Markdown files that encapsulate higher-level meta-skills. This innovation promises to enhance the adaptability and intelligence of AI systems, making them more proficient problem solvers in dynamic environments.