Single AI agents often outperform complex multi-agent systems

New research from Stanford University suggests that single-agent AI systems can match or even surpass multi-agent architectures on complex reasoning tasks when allocated the same computational budget. Multi-agent systems often incur a 'swarm tax' due to increased computational overhead and longer reasoning traces, making it difficult to discern if performance gains stem from architecture or simply higher resource consumption. The study found that single agents are more efficient and cost-effective for multi-hop reasoning, unless the task involves highly degraded contexts where multi-agent systems may offer an advantage. Developers are advised to reserve multi-agent setups for specific bottlenecks rather than as a default.