SkillSmith: Enhancing Local AI Agents via Automatic Skill Construction

A new research paper introduces SkillSmith, a framework designed to improve the performance of locally deployed AI agents by automatically constructing and evolving 'skills'—context-efficient knowledge units. The study, available on arXiv (2608.08037), addresses the limitations of local agents, which use open-source small language models (SLMs) on user-controlled devices, compared to cloud-based agents that rely on closed-source large language models (LLMs). The authors identify that local agents' reduced effectiveness stems from missing environment knowledge, including rules and operation procedures, due to the smaller scale of the backbone models. SkillSmith proposes a cloud-local collaboration where cloud agents help generate skills that are then used by local agents to enhance their task performance without exposing private data or incurring repeated LLM costs. The framework is presented as a solution to supply knowledge non-parametrically and without expert authoring. The paper is authored by researchers affiliated with arXiv, and the announcement type is 'new'. The work is significant for the field of AI agent deployment, offering a potential path to balance privacy, cost, and effectiveness. The paper was published on arXiv under the identifier 2608.08037.