LLM-Based Multi-Agent Systems for Code Generation: A Multi-Vocal Literature Review

A team of researchers led by Zeeshan Rasheeda has published a comprehensive 'Multi-Vocal Literature Review' (MLR) analyzing the burgeoning field of LLM-based multi-agent systems for code generation. By systematically reviewing 114 studies from both academia and industry, the paper provides the first major synthesis of evidence in this rapidly evolving area. The review identifies nine core reasons developers are adopting these multi-agent approaches, which involve multiple AI 'agents' collaborating on complex coding tasks, and systematically catalogs the models and evaluation benchmarks commonly used across the field.

The study's key contribution is a structured analysis of the significant hurdles facing real-world adoption. The researchers synthesized reported challenges and their corresponding solutions into six main categories and 26 subcategories, providing a clear taxonomy of problems like coordination, hallucination, and security. Furthermore, the review outlines six primary directions for future research, broken down into 18 specific subcategories, serving as a vital roadmap. This work is designed to help both academics and industry practitioners navigate the complex landscape and accelerate the development of more robust, deployable AI coding assistants.