New study debunks multi-agent AI advantage over single-agent systems

A research paper titled 'The Illusion of Multi-Agent Advantage' challenges the prevailing belief that Multi-Agent Systems (MAS) are inherently superior to Single-Agent Systems (SAS). The authors, led by Prathyusha Jwalapuram, conducted a rigorous evaluation comparing automatically generated MAS with Chain-of-Thought with Self-Consistency (CoT-SC), a simple but powerful SAS technique. Across traditional reasoning datasets and more complex interactive multi-step workflows like BrowseComp-Plus, the automatic MAS consistently underperformed CoT-SC despite being up to 10x more expensive. The study introduces a diagnostic synthetic dataset designed to test explicit task decomposition, context separation, and parallelization—areas where MAS are claimed to excel. Even on this tailored dataset, expert-architected MAS outperform automated ones, but all MAS still show significant inefficiencies once computational cost is accounted for.

The paper attributes these failures to architectural bloat in current automated MAS design paradigms, which prioritize superficial complexity over functional utility. The automated generation process produces systems that look sophisticated but do not translate into real performance gains, exposing a fundamental misalignment with multi-agent principles. The authors argue that existing evaluation frameworks mask these gaps by failing to consider the marginal utility of increased computational cost. This research has direct implications for AI practitioners: for many tasks, a simpler, single-agent approach like CoT-SC may be more cost-effective and reliable than deploying complex multi-agent architectures.