More agents isn't better: New study reveals diminishing returns in LLM multi-agent systems

A new arXiv paper by Jialing Li, Zhouhong Gu, Yin Cai, and Hongwei Feng systematically investigates how performance changes as the number of agents increases in LLM-driven multi-agent systems (MAS). Their proposed Sequential Iterative Multi-Agent System (SIMAS) framework isolates the effect of agent count by using a homogeneous setup—same LLM for all agents—and sequential inter-agent communication. Through extensive experiments across diverse tasks and model scales, they found that MAS performance does not scale monotonically: there is an initial increase due to collaborative synergy, followed by a plateau and eventual decline due to coordination overhead.

The key insight is that coordination overhead—not merely long-context failures—is the primary culprit behind performance degradation in larger systems. The study also shows that the optimal number of agents depends on both the task type and the base LLM's capability; weaker models benefit less from adding agents. The findings challenge the prevailing assumption that more agents always lead to better outcomes, offering practical guidance for designing efficient collaborative AI systems.