NTT develops foundational technology for Multi-agent AI technology capable of autonomous coordination and supporting business planning tasks

NTT has developed a foundational technology for autonomous coordination among AI agents that, like humans, communicate through dialogue, align their expectations within the team while reading each other's intent, and work together to solve tasks collaboratively. NTT announced this achievement at ACL 2025, one of the most prestigious international conferences in the field of natural language processing.

Multi-agent systems, where multiple AI agents collaborate to handle parts of business operations, have been attracting growing interest. However, existing multi-agent AI systems typically assign divided subtasks to individual agents, making it difficult to maintain consistency across subtasks during task execution.

For example, conventional AI agent systems could not execute complex tasks such as corporate strategy planning, which requires coordination across multiple departments, and multifaceted business planning involving diverse solutions.

In contrast, NTT has developed a new foundational technology for autonomous collaboration among AI agents by enabling them to adopt human-inspired memory structures and co-creative processes. This allows the agents to solve complex tasks much like humans do, by continuously verifying and updating each other's problem-solving approaches and capabilities.

This foundational technology for autonomous collaboration among AI agents that can read each other's intent aims to broadly address business planning tasks that will be essential when AI autonomously manages organizational operations in the future.

NTT is conducting research and development for an AI Constellation. This approach enables AI agents to discuss tasks with and correct each other while offering diverse perspectives, collaboratively creating solutions with humans. The proposed technology was developed as one of the core components of an AI Constellation framework.

With this technology, agents are first generated for each subtask derived from the complex task, and each agent builds knowledge related to its assigned subtask. Human knowledge systems are supported by both individual experiences, known as episodic memory, and generalized facts, known as semantic memory. NTT's technology adopts a mechanism that enables AI agents to accumulate knowledge by combining these two types of memory.

Furthermore, this technology emulates the collaborative creation process found in human society, enabling agents to dynamically acquire and share knowledge with one another. The system also generates expert agents with specialized knowledge necessary for task execution, which participate in discussions to acquire the required knowledge for problem-solving.

To solve complex tasks, agents continuously update their understanding of their own and others' assigned subtasks and approaches through team and production meetings. They regularly align their task-solving strategies and collaboratively integrate diverse subtasks.

This process leads to improvements in execution accuracy and the quality of solutions. NTT's technology enables the generation of high-quality outputs even for complex planning tasks that require consistency, feasibility, and concreteness while satisfying diverse needs, such as formulating integrated corporate branding strategies combining design, public relations, and marketing, as well as developing multifaceted business plans that simultaneously address differing perspectives.

NTT compared the task execution performance of the foundational technology it has developed with conventional methods across various creative document generation tasks. The developed technology achieved results that surpassed conventional methods in both automatic and human evaluation.

For example, in developing business plans reflecting diverse customer needs using tea as a theme, conventional methods simply listed solutions for each agent's subtasks without any combination or mutual complementation of solutions when integrating subtasks.

In contrast, the developed foundational technology has been confirmed to be able to assist the development of a wide range of tea-related products responding to diverse customer needs by integrating detailed information from each agent's tasks and their respective considerations. The foundational technology can also output multifaceted business plans that enhance customer experience through workshops on brewing methods and flavors. Additionally, when evaluating the generated proposals by measuring their alignment with multiple manually prepared reference proposals using ROUGE, the foundational technology showed an average score improvement of approximately 14.4% compared with conventional methods. When reusing knowledge accumulated through previous tasks, the system achieved an average score improvement of approximately 17.2%.

These results demonstrate that the foundational technology developed can generate high-quality output through collaboration among AI agents. Based on these achievements, NTT will advance efforts toward conducting Proof of Concept (PoC) of this autonomous collaboration foundational technology within the current fiscal year.