AMIGO: Agentic Model Instance Generation

Abstract

Generating metamodel-conforming instance models is a recurring task in Model-Driven Engineering (MDE), yet it remains tedious and tool-bound: instances are serialized as verbose, deeply nested XMI that today’s large language models (LLMs) cannot reliably produce or edit. We present AMIGO, a general, harness-agnostic architecture for agentic instance model generation. Instead of letting the LLM emit XMI, an LLM agent expresses modeling intent through a small set of schema-typed query, manage, and validate tools that a metamodel-aware server exposes over the Model Context Protocol (MCP). The server executes every operation deterministically and persists a conforming instance, and the agent uses constraint feedback from instance model validation to self-correct. We demonstrate AMIGO through a ready-to-use instantiation for the Palladio Component Model (PCM): the PCM-MCP server, driven by an LLM agent, turns natural language specifications into a PCM instance that opens in the Palladio Bench. On two microservice systems (Corona-Warn-App and TeaStore), the agent generation produces OCL-valid instances in 596 of 600 runs across three open-weight LLMs, with structures approaching references (per-view Jaccard up to 0.96). AMIGO shows MCP is a viable integration layer between standard agent harnesses and MDE tooling.