Finding the Optimal Balance: NTT Research and Harvard University Study on Agentic AI in the Workplace

In a recent collaboration between NTT Research and Harvard Center for Brain Science, researchers have discovered that the performance of enterprise AI is not solely dependent on the number of AI agents but also on how humans structure, guide, and coordinate multi-agent AI systems.

The research, conducted by Dr. Hidenori Tanaka and Elizabeth Pavlova from NTT Research’s Physics of Artificial Intelligence (PAI) Lab in partnership with Harvard University, challenges the notion that increasing the number of AI agents automatically enhances AI performance. Instead, the study reveals that multi-agent AI systems excel within a specific optimal operating range. Beyond this range, adding more AI agents can lead to conflicting interpretations and hinder collaboration among the agents.

As organizations deploy AI agents across various domains such as customer service, cybersecurity, and scientific research, the question of how to structure AI agent teams for optimal results becomes crucial.

The research findings shed light on how AI organizations function as enterprises adopt agentic AI. It emphasizes that the effectiveness of AI organizations is not solely determined by the quantity of AI agents but also by factors such as communication, organization, and human guidance.

Determining the Optimal Size for Agentic AI Organizations

The study demonstrates that multi-agent AI systems have an optimal operating range where collective performance is maximized. However, beyond this range, additional AI agents can impede effective communication, consensus-building, and problem-solving.

For instance, in the Flag Game experiment, the research revealed that collective accuracy peaked with approximately 16 AI agents before declining. This experiment involved each agent receiving a portion of information about a hidden flag and collaborating to identify the flag’s origin.

The analogy of “too many cooks in the kitchen” is apt, as simply adding more AI agents does not guarantee improved results. Beyond the optimal range, communication challenges arise, leading to diverging viewpoints and decreased collective performance. Thus, the key for enterprise leaders is not to focus on expanding AI organizations but on designing them effectively.

Understanding Multi-Agent AI Behavior

An interactive game like the Flag Game experiment provides valuable insights into how AI agents collaborate, share knowledge, and solve complex problems. It highlights the emergence of communication challenges as AI organizations grow.

Rather than studying individual AI models in isolation, the research integrates principles from physics, mathematics, and machine learning to comprehend how populations of AI agents organize information, communicate, and develop collective intelligence. This approach aligns with NTT Research’s goal of establishing a scientific foundation for trustworthy and scalable AI systems.

Dr. Hidenori Tanaka, the Group Leader of the PAI Lab at NTT Research and the Physics of Intelligence Program at Harvard University, emphasizes the importance of understanding the social dynamics that govern interactions among AI agents. This understanding is crucial as enterprises increasingly adopt multi-agent AI systems.

Effective Structuring of AI Agent Organizations

Key findings regarding agentic AI and enterprise performance include:

• Multi-agent AI systems operate optimally within a specific range. Increasing the number of AI agents may not always enhance collective performance and could even reduce accuracy for certain tasks.
• Organizational design is as important as scale. The communication and collaboration methods of AI agents significantly impact collective performance.
• Diversity in AI models outperforms homogeneity. Teams with a mix of AI models with diverse strengths achieve better results than teams with a single AI model, indicating that diversity can enhance collective intelligence.
• Human guidance shapes effective AI organizations. Clear instructions and communication strategies have a greater impact on collective performance than simply increasing the number of AI agents. Organizations must consider how AI agents communicate, collaborate, and receive guidance from humans for optimal outcomes.

The research underscores the importance of designing effective AI organizations that balance scale, communication, organizational structure, and model diversity. Success in leveraging agentic AI relies not only on the number of AI agents but also on their collaboration, guidance from humans, and complementary capabilities.

The research paper titled “Flag Game: Interpreting Decision Mechanisms of Bounded Social Agents” was presented at the AI4Good Workshop at ICML.

About NTT Research’s Physics of Artificial Intelligence (PAI) Lab

While we have extensive knowledge on building intelligent systems, understanding the underlying principles of intelligence is crucial for the future of AI.

NTT Research’s PAI Lab delves into the fundamental aspects of intelligence to comprehend how intelligent systems learn, reason, communicate, and collaborate. Through research encompassing the physics of AI, neuroscience and AI, AI interpretability, and multi-agent AI systems, the PAI Lab aims to lay the groundwork for trustworthy and scalable artificial intelligence.

About NTT Research

NTT Research, the Silicon Valley research arm of NTT, focuses on pioneering foundational science and translating it into real-world applications across NTT’s global network.

With a team of top scientists across four research pillars, NTT Research advances fields such as optical computing, next-generation cryptography, biodigital twins for precision medicine, and the physics of AI. The organization’s collaboration with NTT Inc., a global enterprise with significant resources dedicated to research and development, positions NTT Research to drive scientific discoveries towards global implementation.

Through initiatives like the Upgrade conference and Scale Academy, NTT Research accelerates the transformation of fundamental research into practical solutions that address real-world challenges across industries.

For media inquiries, please contact:

Cara Milan

AMP Marketing and Public Relations

cara@amppublicrelations.com

415.792.2968