LLMA-UML: Extending the UML for Modeling LLM-Agent Systems

Abstract

Large Language Model (LLM) agents are becoming important building blocks of software systems. However, an analysis of recent publications shows that existing modeling approaches for LLM-agent systems are fragmented and rarely grounded in widely adopted standards such as the Unified Modeling Language (UML). This fragmentation hinders communication, design review, and early architectural validation. After a review of existing modeling approaches, this paper introduces LLMA-UML, a UML profile with semantics for modeling LLM-agent systems. Designed as a lightweight yet semantically explicit modeling extension, it remains compatible with established UML-based workflows and enables model validation via constraints in the Object Constraint Language (OCL). The profile explicitly defines the boundaries of agents, including prompts, responses, context, memory, tools and retrieval. It also separates design-time specifications from runtime instances across UML diagrams. We describe the conceptual foundations and profile design of LLMA-UML and evaluate it through a representative case study. The results indicate that LLMA-UML improves the expressiveness and analyzability of UML models for LLM-based agent architectures.