Auto-scaling LLM-based multi-agent systems through dynamic integration of agents

Abstract

Introduction: Large Language Model-based Multi-Agent Systems (LLM-based MASs) represent a groundbreaking paradigm where diverse LLM-based agents collaborate, leveraging their unique capabilities to achieve shared objectives. Although LLM-based MASs outperform individual agents, their current architectures are limited by predefined, fixed, and static agent designs, restricting adaptability and scalability in dynamic environments. Method: To address these limitations, this study proposes two novel approaches: Initial Automatic Agent Generation (IAAG) and Dynamic Real-Time Agent Generation (DRTAG). These approaches enable the automatic creation and seamless integration of new agents into MASs, driven by evolving conversational and task-specific contexts, thereby reducing the need for human intervention. Our method leverages advanced prompt engineering techniques such as persona pattern prompting, chain prompting, and few-shot prompting to generate new agents through existing LLM agents. Additionally, several evaluation metrics were adapted to score and rank LLM-generated texts. Results: Experimental results demonstrate that the DRTAG approach significantly improves system adaptability and task performance compared to static MAS architectures. The IAAG framework also enhances initial system flexibility, supporting the creation of contextually relevant agents. Discussion: These findings highlight the potential of dynamic LLM-based MASs to overcome the limitations of static architectures to address complex real-world challenges, paving the way for innovative applications across diverse domains.