LLMs for Software Design Exploration: Architect's Augmentation or Replacement?

The exploration of software architecture alternatives is a key part of the architecture design process. In this context, LLM-based techniques — particularly agents — can help architects sift through large design spaces and perform a variety of analyses, extending human capabilities. Despite the impressive developments around LLMs, these techniques face adoption challenges: LLMs often generate convincing decisions and analyses in textual format, which are informative but not always trustable by human architects nor aligned with their expectations. The seminar presents GenAI use cases of LLM-based agents supporting architects in the exploration of pattern-based decisions for requirements, the evaluation of quality-attribute properties of these decisions, and their documentation. The agents rely on RAG, reflection, and agentic workflows. Final discussion on challenges and perspectives for using LLMs in software architecture more effectively, including future lines of research.