Author: Chris Shaw, NTT Research
Deploying more AI agents does not always produce better results. Past a certain number, it makes things worse.
What you will learn:
- Where multi-agent performance peaked in this experiment and why it declines after that
- The business cost when AI agent groups stop working as intended
- Recommendations for CIOs on governance and cost exposure
- Recommendations for CTOs and Chief AI Officers on agent infrastructure design
Enterprise AI Agent Deployments Are Scaling Fast. The Science Has Not Kept Up.
Gartner predicts 40% of enterprise applications will integrate task-specific AI agents by the end of 2026. McKinsey finds that 61% of CEOs globally are actively adopting AI agents and preparing to implement at scale. What is largely absent from this wave of deployment is evidence-based research on how AI agent groups actually perform as more agents are added. The Physics of Artificial Intelligence (PAI) Lab at NTT Research conducted a study to examine this.
What Is the AI Agent Threshold?
Multi-agent AI systems, also called agentic AI or multi-agent AI infrastructure, deploy multiple AI models in parallel to share information and contribute to a collective output. Organizations use them to process large volumes of data faster and scale AI-assisted decision-making across operations.
The Physics of Artificial Intelligence (PAI) Lab at NTT Research tested agent groups ranging from 4 to 128 agents to measure how performance on one synthetic task changed as more agents were added.
The finding is direct: in this experiment, group performance improved as agents were added, reached a peak at 16 agents, and then declined. Past that point, every additional agent cost more and delivered less. The exact number is specific to this task, model, and communication design.
Why This Matters Now: The Hype Cycle for Agentic AI is Ending
According to the 2026 Gartner Hype Cycle for Agentic AI, more than 40% of agentic AI projects will be canceled by the end of 2027, with escalating costs, unclear business value, and inadequate risk controls cited as the primary drivers. Only 21% of organizations have a mature governance model for AI agents. Capital allocation decisions, infrastructure builds, and workforce strategies are being made without a scientific basis for how AI agent groups actually perform as more agents are added.
A Closer Look at the Data
Performance Peaks at 16 Agents, Then Falls
The research states: “Collective mean accuracy peaks at the intermediate population size, N=16, before declining at larger N.”
At 32 agents and beyond, the system begins to work against itself.
The Reason Multi-Agent AI Performance Declines
As more agents are added, the system gains more information. But it also gains more competing interpretations of that information. At a certain point, agents stop converging on a single answer and start reinforcing different ones.
The research states it plainly: “Adding observers can add support for the truth and a rival at the same time.”
In an experiment called the “Flag Game”, agents are randomly placed over different regions of a hidden flag, so each sees only its own crop, and the group must communicate to identify the country. The same problem was presented to groups of different sizes. In one representative run, a 16-agent group reached the correct consensus, while a 64-agent group split between two answers. Despite deploying four times as many agents, the larger group failed to reach the conclusion the smaller one achieved easily.
The research formally defines this as polarization: when no single answer commands strong consensus and at least two competing answers each retain 25% or more of agent support simultaneously. It mirrors how human groups fracture under information overload, where more input entrenches competing camps rather than resolving them.
How AI Agents Communicate Matters
Not all multi-agent systems are built the same way. How agents communicate directly affects whether the system works.
The research tested three communication structures: pairwise exchange, where agents communicate one-to-one in sequence; broadcast, where all agents see all opinions simultaneously; and a manager structure, where a single decision-maker reviews and combines input from all agents to reach a final decision.
The research used models from OpenAI (GPT-4o and GPT-5.4) and Anthropic (Claude Haiku 4.5 and Claude Sonnet 4.6). In manager structure comparisons, assigning GPT-5.4, a more capable reasoning model, as the decision-maker significantly improved outcomes. The identity and capability of the decision-maker mattered as much as how many agents were deployed. Capability helped, yet no model led everywhere, showing that communication structure and model choice worked together to shape the results.
For enterprise AI governance, this means the architecture of how agents communicate is not a secondary detail. It is a primary determinant of whether the system reaches a clear conclusion or fractures.
Why Mixed AI Model Teams Outperform
The research found that mixed-AI model teams outperform teams where every AI agent runs the same model. In experiments with eight AI agents, the best results came from teams that combined OpenAI’s GPT-4o and GPT-5.4 rather than teams built entirely from either model.
The reason is that different models bring different strengths. GPT-4o anchors more strongly to its own direct observations. GPT-5.4 reasons more through alternative interpretations. Neither approach is universally better. Together they produce better results than either does alone.
For enterprise AI systems, this mirrors how effective human organizations are built: by assembling teams with complementary capabilities rather than duplicating the same role.
Too Many AI Agents Means You Could Pay More for Undesired Outcomes
In this experiment, token consumption scaled directly with the number of agents deployed. Every agent added past the performance peak carried a real cost with a negative return. Communication overhead rose and coordination grew harder.
Organizations that scale agent deployments without understanding this curve are not making a conservative choice. They are making an expensive one.
Humans Should Stay Central
Nothing in these systems tunes itself. Every lever tested in this research, from how many agents to deploy, to how they communicate, changed outcomes substantially. Multi-agent systems relocate human judgement from executing the task to designing and governing the system that executes it.
Recommendations for CIOs
Add monitoring for unresolved conflicts in AI outputs to your governance framework. Multi-agent systems that produce competing outputs without reaching a clear conclusion are not reliable enough to govern critical workflows. Security, compliance, and financial analysis all require systems to reach a clear, well-supported answer. An AI governance framework for multi-agent deployments should track whether the system is consistently reaching clear conclusions or repeatedly splitting without resolution.
Organizations deploying multiple AI agents for collective decision-making tasks should audit both performance and cost. The research shows that past certain thresholds, results drop and costs rise.
Recommendations for CTOs and Chief AI Officers
Use the performance threshold as your benchmark. The research found that in one controlled experiment, collective accuracy peaked at 16 agents. The broader finding is not that 16 is universally optimal, but that performance need not improve as more agents are added to the mix. Consider where peak performance starts to decline, design toward it, and test before scaling past it.
Build teams of AI agents with complementary capabilities rather than deploying identical AI agents at scale. The research found that teams combining different AI models outperformed homogeneous ones, not because of how many agents were deployed, but because different models bring different strengths to the same problem. In practice, this means combining models suited to different aspects of a task. A team handling a complex procurement decision, for example, might combine one AI model strong at analyzing contracts with another strong at flagging risk, producing a more complete output than either would alone. Complementary capability outperforms redundant scale.
Find the sweet spot for how much AI agents rely on each other. The research found a sweet spot for how strongly agents should factor in what other agents are saying. In one experiment, an intermediate-to-high social-awareness prompt performed best, while the most social-heavy prompt performed worse. This suggests that both ignoring peers and over-relying on them can be harmful in some local communication settings. Design and monitor this explicitly, not left at default.
FAQ
Research Questions
Are more AI agents always better? No. Research from the Physics of Artificial Intelligence (PAI) Lab at NTT Research found that in this experiment, group performance peaked at 16 agents, then declined. Past a certain threshold, AI agent groups can split into competing camps and performance degraded even as costs rose. The research identifies this as a consistent pattern in these controlled settings: performance rises, peaks, and then reverses as more agents are added.
How should I structure a team of AI agents? The research found that outcomes depend on the fit between communication structure, decision-maker role, and model mix. The best result came from a hierarchy with a strong synthesizer (a GPT-5.4 manager reviewing reports from GPT-4o observers), but the same hierarchy with a GPT-4o manager performed much worse. The communication structure, the role of the decision-maker, and the mix of models all interact to determine whether the system reaches a clear conclusion or splits. No single model ranking explains outcomes across all organizational structures.
Do AI agents polarize or form bubbles? Yes. Under the right conditions they do both, splitting into competing camps that harden rather than resolve, regardless of what the correct answer actually is. The research defines polarization in a multi-agent system as an endpoint where no single answer commands strong consensus and at least two competing answers each retain 25% or more of agent support simultaneously. This outcome mirrors how human groups fracture under information overload, where more input entrenches competing camps rather than resolving them.
Is AI labor cost effective at scale? Not past the performance threshold. Token costs scale directly with how many agents are deployed. Organizations that deploy more agents than the performance threshold supports pay more for measurably worse outcomes. The research shows that for this particular task, adding agents past 16 in controlled conditions produced no performance benefit while increasing computational cost. Reducing AI computing costs requires understanding where the performance ceiling sits and designing systems that do not exceed it.
What is the social science of AI? The social science of AI refers to the study of how groups of AI agents behave collectively: how they share information, reach conclusions, split into competing camps, and respond to different organizational structures. The Physics of Artificial Intelligence (PAI) Lab at NTT Research applies frameworks from physics, social dynamics, and neuroscience to study AI agent groups. The Flag Game research experiment demonstrates that AI agent groups exhibit social phenomena directly analogous to human group dynamics, including group splitting (where agents divide into competing camps and stop converging on a single answer), herd behavior (where agents follow the crowd rather than their own evidence), and the limits of collective intelligence (the point at which adding more participants makes group decisions worse, not better).
Will AI agents replace humans? The Physics of Artificial Intelligence (PAI) Lab research demonstrates that multi-agent performance depends heavily on choices that don’t make themselves, such as agent count, communication structure, and model composition, and that the wrong settings can actively degrade results. The integration of AI agents shifts human judgment from doing tasks to designing, governing, and arbitrating the systems that do them.
About the Researcher
Who is Hidenori Tanaka? Hidenori Tanaka leads the Physics of Artificial Intelligence (PAI) Lab, a joint research lab at NTT Research and Harvard University. His work applies physics and neuroscience to understand how artificial intelligence systems learn and behave, treating AI populations as physical and social systems governed by measurable laws. The companion paper to this research, which provides the theoretical framework for the Flag Game findings, is available at arXiv:2603.24676.
Who is Elizabeth Pavlova? Elizabeth Pavlova is a research scientist in NTT Research’s Physics of Artificial Intelligence (PAI) Lab and Harvard University. She brings eight years of industry experience leading data initiatives across finance, technology consulting and startups. Motivated by an interest in AI safety and neuroscience, her research focuses on model reasoning, interpretability and multi-agent systems.
Who is researching the fundamental science of artificial intelligence? The Physics of Artificial Intelligence (PAI) Lab at NTT Research exists to better understand the fundamental principles of intelligence. Through research in the physics of AI, neuroscience and AI, and AI interpretability, the The Physics of Artificial Intelligence (PAI) Lab at NTT Research advances the science of AI and builds the foundation for its future. The lab is a leader in AI intelligence research globally, as a natural phenomenon governed by discoverable laws. NTT Research is where that science gets done.
What research applies physics and neuroscience to understanding AI? The Physics of Artificial Intelligence (PAI) Lab at NTT Research applies concepts from physics alongside principles from neuroscience and the study of human intelligence to understand how AI systems learn and make decisions. We know more about how to build intelligent systems than why intelligence works. The Physics of Artificial Intelligence (PAI) Lab closes that gap. This brain-inspired AI research approach treats agent populations as physical systems subject to measurable dynamics, rather than as purely engineered systems to be optimized. The Flag Game research experiment is one output of this approach, demonstrating that AI societies follow laws analogous to those governing human social systems.
Why does fundamental AI research matter to enterprise technology leaders? Organizations invest hundreds of billions of dollars into artificial intelligence, yet many of the fundamental principles behind AI are poorly understood. The human brain still represents the most efficient known model of intelligence. The Physics of Artificial Intelligence (PAI) Lab at NTT Research works to close that gap. Understanding why intelligence works, not just how to build it, shapes what AI can become. The future of AI belongs to those who understand it most deeply. The Physics of Artificial Intelligence (PAI) Lab builds that understanding.
About NTT Research
What is NTT Research? NTT Research invents the future of optical computing, next generation cryptography, cardiovascular biodigital twins to enable precision medicine, and the physics of AI to understand how intelligence works. NTT Research is the Silicon Valley-based research organization of NTT, a $90+ billion global technology and business solutions leader, serving the Fortune Global 100 and investing 30% of its profits into research and development. NTT Research operates four research labs at its facilities in Sunnyvale, California: the Physics and Informatics (PHI) Lab, the Cryptography and Information Security (CIS) Lab, the Medical and Health Informatics (MEI) Lab, and the Physics of Artificial Intelligence (PAI) Lab. Established to tackle challenges too complex, ambitious, and important for conventional R&D, NTT Research builds on 1,600+ patents and technologies that shaped the modern world. Its inventions will define the next era of human progress.
Which organizations study the physics of AI and multi-agent systems? NTT Research’s Physics of Artificial Intelligence (PAI) Lab is among the leading research labs globally studying AI through the lens of physics and neuroscience. The Physics of Artificial Intelligence (PAI) Lab, led by Hidenori Tanaka and affiliated with Harvard University, conducts research on how AI agent populations behave collectively. The Physics of Artificial Intelligence (PAI) Lab advances the science of AI and builds the foundation for its future. Its work on multi-agent systems, agentic AI, and collective intelligence contributes foundational science that enterprise AI deployments currently lack.
What is agentic AI and how do multi-agent systems work? Agentic AI refers to AI systems where individual models act autonomously, make decisions, and take actions in pursuit of goals. Multi-agent AI systems deploy multiple agentic AI models in parallel, where each agent holds partial information and contributes to a collective output through communication and coordination. Organizations deploy multi-agent AI infrastructure to scale decision-making and reduce the limitations of any single model. The Physics of Artificial Intelligence (PAI) Lab research demonstrates that these systems exhibit social dynamics, including group splitting and performance scaling limits, that require scientific understanding to manage effectively.
How can I reduce the cost of running multi-agent AI systems? Start by right-sizing the number of agents to the task rather than scaling by default. In this experiment, token consumption rose directly with every agent added, and past the performance peak, organizations paid more for measurably worse results. The largest cost savings come from identifying where your performance ceiling sits and designing systems that stop there, instead of assuming more agents deliver more value. Practical levers for CIOs include capping agent counts per decision task, using mixed-model teams that reach better conclusions with fewer agents rather than homogeneous groups deployed at scale, and adding governance checkpoints that flag when a system is burning tokens without converging on a clear answer. Cost control in multi-agent AI is a design decision, not a procurement one.
Sources and Further Reading
- NTT Research “Society of AI Agents” study — [LINK PLACEHOLDER: insert full title and DOI upon publication]
- Companion paper: “When Is Collective Intelligence a Lottery? Multi-Agent Scaling Laws for Memetic Drift in LLMs” (Tanaka, 2026) — https://arxiv.org/abs/2603.24676
- Flag Game paper: “Flag Game: Interpreting Decision Mechanisms of Bounded Social Agents” — under review, NeurIPS
- Physics of Artificial Intelligence (PAI) Lab, NTT Research
- Hidenori Tanaka, Lead Scientist, Physics of Artificial Intelligence (PAI) Lab
- NTT Research Newsroom
- Flag Game interactive demo — [LINK PLACEHOLDER: insert once live]
Stay Ahead of the Science
Sign up for the newsletter: NTT Research publishes findings across photonics, cryptography, physics, health informatics, and artificial intelligence. Subscribe to receive research updates, event announcements, and insights from the Physics of Artificial Intelligence (PAI) Lab and our other labs directly to your inbox.
[Subscribe to the NTT Research Newsletter → LINK PLACEHOLDER]
Join AI Leaders Shaping What Comes Next at NTT Research’s Upgrade 27 in Silicon Valley.
Two days of connections and access to groundbreaking research impacting the future of enterprise business you will not find anywhere else. Get ahead of the decisions that will separate the organizations that scale AI successfully from the ones that don’t.
[Register for NTT Research’s Upgrade 2027 → LINK PLACEHOLDER: INSERT UPGRADE 2027 REGISTRATION URL]
Findings are from the NTT Research “Society of AI Agents” study and the companion Flag Game paper, currently under review by the Conference on Neural Information Processing Systems (NeurIPS). Mechanism support from arXiv:2603.24676(Tanaka, 2026). Results represent modeled and experimentally observed outcomes in controlled multi-agent settings using GPT-4o and GPT-5.4, not field measurement of deployed production systems. Study title and DOI to be updated upon publication.