LLM-Based Multi-Agent Systems for Software Engineering: Literature Review, Vision, and the Road Ahead

Single-agent LLM pipelines plateau on software engineering tasks that need division of labour: planning, delegating, reviewing, and revising. Multi-agent (LMA) systems promise exactly that decomposition, and the literature grew quickly enough that its shared assumptions went unaudited.

This paper reviews that literature, organizes it around the SE problems it targets, and runs two case studies with state-of-the-art LMA frameworks to see how the promise holds up in practice. We then lay out the open challenges — coordination cost, error propagation between agents, and the absence of evaluation that isolates the contribution of multi-agency itself — along with a research roadmap.