From User Stories to Architectures: Using LLM-powered Agents to Design and Improve Microservice-based Software

Large Language Models have rapidly transformed how developers write and test code, yet their use in software architecture design remains limited. Current AI-powered assistants can generate useful snippets of code but lack a holistic view of how modules should interact to form reliable systems. We discuss ArchiLLM, an LLM-based framework that extends the capabilities of AI assistants from code generation to architecture synthesis. Given minimal textual input, such as requirements or user stories, ArchiLLM envisions a high-level microservice-based architecture before generating code, aiming to improve the reliability and coherence of AI-produced systems. The approach is validated using the Archi Dataset, built from academic and open-source microservice projects, and evaluated through quantitative metrics and expert feedback. Results suggest that ArchiLLM can effectively assist software designers in moving from abstract requirements to consistent architectures to better code, marking a step toward AI-supported software design and development.