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

Researchers present a formal architectural framework for managing LLM agent skills—reusable behavioral components that agents dynamically select and execute. The paper catalogs ten architectural patterns organized into four responsibility layers (Supply Chain, Mediation, Execution Control, Evidence & Feedback) and provides a reference architecture validated across eight systems, establishing a standardized approach for skill governance in agent-based AI applications.

This academic work addresses a critical gap in LLM agent system design by formalizing how reusable behavioral components transition from static artifacts to active, context-specific execution. The research distinguishes between skill-at-rest (a persistent definition) and skill-in-use (runtime binding with authority constraints and stochastic interpretation), a distinction that becomes essential as agent systems grow in complexity and operational scope.

The framework emerges from practical challenges in deploying autonomous agents at scale. Organizations deploying multi-agent systems face fragmented approaches to skill discovery, activation, and accountability. The paper synthesizes these concerns into architectural patterns addressing supply chain (skill provisioning), mediation (selection and binding), execution control (runtime guardrails), and evidence collection (auditability). This layered approach parallels DevOps practices in software engineering, suggesting the maturation of agent infrastructure toward production-grade standards.

The impact extends across AI development, particularly for enterprises building internal agent platforms. Developers gain a diagnostic vocabulary for evaluating architectural gaps in their systems, while organizations gain frameworks for implementing reproducible, auditable agent behaviors. The cross-system validation across eight implementations indicates the patterns' generalizability beyond specific frameworks.

Future developments likely involve standardization of skill definition formats and governance protocols. As regulatory scrutiny on AI systems intensifies, the evidence-and-feedback layer's emphasis on attribution and verification becomes increasingly valuable. Teams building agent infrastructure should monitor these architectural developments to ensure compatibility with emerging compliance and safety standards.

• →The paper formalizes skill-in-use as a distinct architectural concern separate from static skill definitions, requiring governance across four responsibility layers.
• →Ten empirically grounded patterns provide a reusable vocabulary for designing skill-harnessing systems in LLM agents.
• →The framework's cross-system validation indicates applicability across diverse agent architectures and implementations.
• →Evidence-and-feedback mechanisms enable attribution, verification, and repair—critical for auditable agent systems.
• →The architecture patterns align with enterprise DevOps practices, suggesting agent systems are entering production maturity phases.