Title: Scaling Dynamics in Single LLM-Powered Multi-Agent Frameworks

Abstract:

While Large Language Model (LLM)-based Multi-Agent Systems (MAS) hold significant potential for addressing intricate challenges through shared intelligence, their scaling characteristics and underlying collective behaviors remain largely unexamined. This study systematically analyzes how the efficacy of a homogeneous MAS changes as the agent population grows, specifically isolating collaboration as the primary variable while controlling for variations in model capabilities or knowledge bases. To clearly isolate scaling effects, we introduce the Sequential Iterative Multi-Agent System (SIMAS), a streamlined architectural framework that relies on sequential communication channels between agents.

Our comprehensive experiments, conducted across a wide range of tasks and model sizes, demonstrate that MAS performance does not increase linearly with the number of agents. Instead, the system exhibits a pattern of diminishing returns, driven by a tension between the benefits of collaborative synergy and the costs associated with coordination overhead. The results indicate that successful MAS implementation depends on having a sufficiently robust base LLM, while the ideal number of agents is heavily influenced by the specific nature of the task at hand. Furthermore, we find that collective intelligence is an emergent phenomenon resulting from strategic interaction design, rather than an automatic consequence of having multiple agents. Notably, performance declines are attributed to coordination burdens rather than simply limitations in long-context processing. These scaling trends appear to be consistent across various interaction structures, including structured debate topologies. By establishing foundational insights into MAS scaling laws, this research offers practical recommendations for building efficient collaborative systems and challenges the common misconception that increasing the number of agents always yields superior outcomes.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC