Harnessing Agent Skills: Architectural Patterns and a Reference Architecture for Skill-Mediated LLM Agents

The paper presents a reference architecture for skill-mediated LLM agents, detailing ten architectural patterns that facilitate the transition from static skill artefacts to dynamic skill-in-use. It introduces a framework comprising four responsibility layers: Supply Chain, Mediation, Execution Control, and Evidence & Feedback, evaluated through cross-instantiation across eight systems. This work is significant for practitioners as it offers a structured approach to integrating reusable agent skills, enhancing the robustness and adaptability of LLM applications.