Google-MIT Study: Multi-Agent AI Can Cut Performance by 70% Without Right Design

A joint study by Google Research and MIT tested 180 configurations of multi-agent AI systems across three model families and five network architectures, keeping prompts, tools, and compute budgets identical while only changing how agents were connected. Researchers found that performance swung dramatically — from an 81 percent improvement to a 70 percent drop — depending solely on the network topology used. For parallelizable tasks, a centralized coordination structure outperformed a single agent by over 80 percent, while sequential reasoning tasks saw every multi-agent variant degrade performance by 39 to 70 percent. Averaged across all experiments, multi-agent systems delivered virtually no overall gain compared to a single agent, at just plus 0.2 percent. The findings suggest that adding more AI agents without deliberate architectural planning can be counterproductive, and that connection design matters more than model choice or agent count.