https://www.mdu.se/

LLM–based multi-agent systems (LMAS) are emerging as a promising paradigm, enabling specialized agents to collaborate and act autonomously across complex software engineering (SE) tasks. Yet, despite rapid progress, the field still lacks a clear conceptual foundation to guide researchers and practitioners in systematically designing, implementing, and evaluating such systems. This article introduces a taxonomy intended to provide practical guidance for integrating LMAS into software development workflows. Grounded in an analysis of prominent frameworks and scientific contributions for engineering LMAS, the taxonomy organizes the design space along five key dimensions offering a structured lens for understanding and engineering LMAS. Our analysis reveals substantial gaps in current frameworks, particularly regarding monitoring, performance evaluation, and quality assessment. These findings highlight pressing research challenges and identify concrete directions toward more reliable, transparent, and effective agentic systems for SE.