The Illusion of Multi-Agent Advantage

This paper presents a critical evaluation of Multi-Agent Systems (MAS) compared to Single-Agent Systems (SAS), challenging the assumption that MAS inherently offer superior performance. The authors demonstrate that automatically generated MAS, despite being up to 10x more expensive, consistently underperform against the Chain-of-Thought with Self-Consistency (CoT-SC) method on various reasoning tasks, particularly in interactive workflows. The findings highlight the inefficiencies and architectural bloat in current MAS design paradigms, suggesting that existing evaluation frameworks may overlook key performance metrics, which is crucial for practitioners aiming to optimize multi-agent architectures in real-world applications.