Can LLM agents with generic tools perform MBSE tasks, build-up skills and transfer them?

Abstract

A Model-Based Systems Engineering (MBSE) model can be construed as a complex graph database, and this is true regardless of the underlying modeling language (SysML, SDL, Capella or indeed any Domain Specific Modeling Language). It is hence conceivable that an intelligent artificial agent, for example one based on a Large Language Model (LLM), may perform model-based tasks like creating or editing model elements using only generic CRUD (Create/Read/Update/Delete) tools, provided that these tools are made available through the Model Context Protocol (MCP), a tool-access protocol that most agents support. It is however not obvious that the agent can be effective at performing those model-based tasks, as MBSE models have to satisfy many internal coherence rules, some (or many) of which are not explicitly specified in language or tool manuals. This paper presents an experiment in building an MCP server for a complex MBSE environment (Capella) based on generic model access functions of the Eclipse Modeling Framework (EMF), evaluating the effectiveness of agents at learning how to use the MCP tools to perform a range of model editing tasks in Capella, and finally, evaluating the effectiveness of in-context transfer of the learned skills to other non-skilled agents.