For enterprise agents, less is more
NTT and Harvard researchers found optimal multi-agent performance at 16 agents; scaling beyond triggers polarization and consensus breakdown. For tech leaders: agent orchestration ROI plateaus—prioritize quality and coordination over scale quantity.
NTT Research and Harvard show multi-agent AI peaks around 16 agents; additional agents reduce accuracy as communication fractures consensus. Enterprise implication: agentic AI effectiveness depends on organizational design and human guidance, not agent count.
How Many AI Agents Are Too Many? NTT Research and Harvard University Research Reveals Key Insights for Using Agentic AI in the Workplace