As companies race to deploy AI agents across customer service, coding, cybersecurity, and research workflows, a new question is emerging: is there such a thing as too many agents?
According to new research from/NTT Research’s Physics of Artificial Intelligence (PAI) Lab in cooperation with Harvard University’s Center for Brain Science, the answer is yes.
The study finds that multi-agent AI systems have an optimal operating range — and that piling on more agents past that point can actually hurt performance rather than help it.
The team, led by Dr. Hidenori Tanaka alongside Data Scientist Elizabeth Pavlova, tested this premise through an experiment entitled the “Flag Game.” Each AI agent is given only a small, randomly assigned piece of a hidden flag, and the group has to communicate to figure out which country’s flag it represents. It’s a clean way to study how distributed knowledge, communication, and consensus-building play out at scale.
The results: collective accuracy peaked at a certain point . Add more agents beyond that, and communication starts to break down — competing interpretations emerge, groups splinter into camps, and accuracy drops. As Tanaka put it, it’s a bit like how hiring more people doesn’t automatically make a company more effective.
A few other findings stand out for enterprise leaders building agentic AI systems:
- Structure matters as much as scale.
How agents communicate and exchange information has a bigger impact on outcomes than simply how many of them there are.
- Diversity beats uniformity.
Teams mixing AI models with complementary skills outperformed teams built from a single model.
- Human guidance still counts.
Explicit instructions and communication tactics from human overseers moved the needle more than adding headcount — or agent-count — alone.
The upshot for organizations experimenting with multi-agent deployments: the goal isn’t to build the biggest possible AI workforce, but the best-organized one.
As enterprises move from single-agent tools toward systems with dozens, hundreds, or even thousands of agents working together, questions of coordination, model diversity, and human monitoring are likely to matter more than raw scale.
The paper, “Flag Game: Interpreting Decision Mechanisms of Bounded Social Agents,” was accepted for presentation at the AI4Good Workshop at ICML.
It’s part of NTT Research’s broader push to understand the fundamental principles behind how intelligent systems — human or artificial — learn, reason, and collaborate.
