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

Analysis

Why This Paper Matters

Large Language Models have rapidly moved from code completion to autonomous problem-solving, and multi-agent architectures are emerging as a way to tackle the complexity of real-world software projects. This paper arrives at a pivotal moment: instead of proposing yet another single-agent tool, it steps back to systematically map how LLM-based multi-agent (LMA) systems are being applied across the software development lifecycle (SDLC). For practitioners and researchers alike, this provides a much-needed organizing framework for a fast-fragmenting field.

The significance is amplified by the paper's venue and timing. Published in ACM Transactions on Software Engineering and Methodology in January 2025, it bridges the software engineering and AI communities, offering a shared vocabulary for agent roles, collaboration patterns, and lifecycle coverage. With 411 citations already, it is clearly being treated as a reference point for the emerging subfield of agentic software engineering.

Moreover, the paper does not stop at description. By pairing the review with two case studies and a research agenda, it connects current practice to future directions, making it useful both as a survey and as a roadmap.

Technical Contributions

• Systematic literature review: A structured mapping of recent primary studies on LMA systems across SDLC stages, from requirements to maintenance.
• Case studies: Two demonstrations of state-of-the-art LMA frameworks that illustrate current capabilities and expose practical limitations.
• Research gap analysis: Identification of critical gaps at two levels—individual agent capabilities (planning, reasoning, robustness) and agent synergy (collaboration, coordination, trust).
• Research agenda and vision: A forward-looking agenda for fully autonomous, scalable, and trustworthy LMA systems, framed as the evolution toward Software Engineering 2.0.

Results

The paper's results are primarily qualitative and synthetic rather than benchmark-driven. The systematic review maps the landscape of LMA applications across the SDLC, showing where multi-agent approaches have been applied and where coverage remains thin. The two case studies demonstrate the effectiveness of state-of-the-art frameworks in practice, while also surfacing limitations that motivate the proposed research agenda. No standardized quantitative metrics or head-to-head comparisons are reported in the abstract; the contribution is the synthesis, the illustrative evidence, and the agenda rather than a new benchmark or model.

Significance

The broader impact of this work lies in shaping how the AI and software engineering communities think about agentic systems. By framing LMA systems as a path toward autonomous, scalable, and trustworthy software engineering, the paper sets an agenda that spans agent-level improvements and multi-agent coordination. Its vision of Software Engineering 2.0 positions multi-agent LLM systems not as isolated tools but as collaborative infrastructures for managing real-world software complexity. As citations accumulate, the paper is likely to serve as a foundational reference for researchers designing agent architectures and for practitioners evaluating where multi-agent approaches add value across the development lifecycle.