A new arXiv paper, "Scaling Multi-Agent Intelligence with Substrate-Centric Coordination" (arXiv:2610.16120), examines why large language model (LLM) based multi-agent systems struggle to scale. The authors identify three problems that emerge as the number of agents increases, though the abstract as provided details only two of them.

The first problem is a reading burden that grows with the group: each agent must process more and more information from its peers. The second is that useful information becomes diluted by shared work of unknown quality, making it harder for agents to find what matters. The third problem is mentioned but not specified in the available abstract.

The paper's title suggests a "substrate-centric" approach to coordination, but the source does not describe the proposed solution. The findings point to fundamental bottlenecks in current multi-agent designs, which may inform future work on scaling collective LLM intelligence.