Multi-agent systems built from large language models are often deployed on the assumption that collaboration beats a single agent. But a new arXiv preprint asks when that assumption actually holds, testing agent scaling in what it calls dual-control settings. The abstract describes it as the most systematic study to date on the question.
According to the abstract, the key result is a capability-saturation effect: the advantage of adding more agents appears to level off as the capability of individual agents increases. In other words, collaboration may pay when agents are weak, but the benefit may shrink once agents become strong enough on their own.
The paper does not yet provide full details in the abstract, and the exact conditions under which saturation occurs are not specified here. Still, the finding adds nuance to the growing debate about whether multi-agent LLM systems are worth their complexity.