LLM-Powered Multi-Agent Systems: Exploring Documentation-Driven Metamodeling

By: James Pontes Miranda, Ansgar Radermacher, Fabien Baligand, Julie Bonnail, Kunal Suri, Pascal Bannerot, Marcos Didonet Del Fabro

Abstract

In Model-Driven Engineering (MDE), metamodeling is a crucial activity and is often the starting point of a full MDE pipeline. A central aspect of this task is extracting domain knowledge from textual documentation and establishing the necessary classes, relationships, and constraints that will be later used to formalize the metamodel. This early stage is known to be demanding, error-prone, and influenced by individual modelers’ expertise and bias. Recent advancements in the use of Large Language Models (LLMs), have stimulated research on interpreting textual information to generate models and metamodels. However, there is still limited work exploring the potential of LLM-powered agents, particularly LLM-powered Multi-Agent Systems (LLM-MAS), to support the metamodeling process in a structured manner. In this paper, we present an approach that leverages an LLM-MAS to assist in documentation-driven metamodeling. The proposed approach decomposes the task into multiple specialized agents responsible for activities such as domain analysis, terminology identification, normalization and deduplication, and textual serialization into a PlantUML class diagram. The system operates without human intervention and produces intermediate artifacts that support traceability and inspection. We report on the development of this LLM-MAS and its application in a case study involving the extraction of a draft metamodel from a set of agent framework documentations. We provide an exploratory qualitative evaluation focusing on the feasibility, stability, and structural plausibility of the generated artifacts. The results indicate that LLM-MAS can consistently produce structurally plausible model-like artifacts that may assist modelers in the early stages of metamodel creation. Rather than targeting full automation, the approach positions LLM-MAS as a modeling aid that supports early abstraction and helps modelers initiate metamodel development more systematically.

Keywords

Metamodeling, Large Language Models, LLM4MDE, Multi-Agent Systems

Cite as:

James Pontes Miranda, Ansgar Radermacher, Fabien Baligand, Julie Bonnail, Kunal Suri, Pascal Bannerot, Marcos Didonet Del Fabro, “LLM-Powered Multi-Agent Systems: Exploring Documentation-Driven Metamodeling”, Journal of Object Technology, Volume 25, no. 3 ( 2026), pp. 3:379-392, doi:10.5381/jot.2026.25.3.a29.