Auto-Scaling LLM-Based Multi-Agent Systems through Dynamic Integration of Agents

Ravindu Ramesh Perera^*Anuradha BasnayakeManjusri Wickramasinghe

• University of Colombo, Colombo, Sri Lanka

Large Language Model-based Multi-Agent Systems represent a groundbreaking paradigm where diverse LLM-based agents collaborate, leveraging their unique capabilities to achieve shared objectives. Although LLM-based MASs have shown significant improvements over individual agents, their current architectures are constrained by predefined, fixed, and static agent designs with limited adaptability and scalability in dynamic environments. To address these limitations, this study proposes two novel approaches. Initial Automatic Agent Generation (IAAG) and Dynamic Real Time Agent Generation (DRTAG). These frameworks enable the automatic creation and seamless integration of new agents into MASs in response to evolving conversational and task-specific contexts, reducing reliance on human intervention. Key contributions of this research include the development of IAAG and DRTAG for automatic agent integration, adaptation of several evaluation matrices to score and rank LLM generated texts, and advanced prompt engineering techniques, such as persona pattern prompting, chain prompting, and few-shot prompting, to facilitate the creation of new agents via existing LLM agents. Experimental results demonstrate that the DRTAG approach, particularly with the Round-Robin agent selection algorithm, significantly enhances the system adaptability and task performance. These findings highlight the potential of dynamic LLMbased MASs to address complex real-world challenges, paving the way for innovative applications across diverse domains.