In this paper, the authors present MetaGPT, an innovative framework that addresses the limitations of existing language model-driven multi-agent systems. These systems have been successful in solving simple dialogue tasks but struggled with complex problems due to the “hallucination problem.” MetaGPT incorporates efficient human workflows and encodes Standardized Operating Procedures (SOPs) into prompts to enhance structured coordination. It also mandates modular outputs and leverages human domain expertise to validate and minimize errors. The authors conducted experiments on collaborative software engineering benchmarks, demonstrating that MetaGPT outperforms existing systems in generating coherent and correct solutions. The framework shows promise in tackling complex real-world challenges by integrating human domain knowledge into multi-agent systems.
https://arxiv.org/abs/2308.00352