SkillSmith Boosts Local AI Agents with Automatic Skill Learning

Large Language Model (LLM) based agents are increasingly used as personal assistants for complex tasks. While cloud-based agents leverage powerful, often proprietary LLMs, they raise concerns about data privacy and recurring costs. Local agents, using smaller, open-source models on user devices, address these issues but typically lag in performance due to a lack of environmental knowledge. A new framework, SkillSmith, tackles this by enabling a collaboration between cloud and local agents. It automatically constructs "skills" – context-efficient knowledge carriers – by observing cloud agent task explorations. These skills are then refined and evolved through feedback from the local agent's own execution, effectively enhancing the local model without requiring expert manual authoring. Experiments demonstrate that SkillSmith significantly improves the effectiveness of local agents, allowing them to perform multi-step tasks with accuracy comparable to their cloud-based counterparts. For instance, it reduced the average actions per task on AppWorld-Normal from 36.1 to 9.9, and the approach generalizes to different local models.