Fujitsu Begins Early Validation of Fujitsu Kozuchi Multi AI Agent Framework Incorporating Self-Evolving Multi-AI-Agent Technology
From AI Agents That Are "Built and Finished" to AI Agents That Continue Evolving With Business Operations

In recent years, efforts to apply AI agents to business operations have accelerated rapidly. In real-world workplaces, however, business procedures, exception handling, decision criteria, and system operations are intricately intertwined. As a result, even after AI agents are built, they often cannot keep pace with regulatory revisions, specification changes, or shifting customer needs, causing initiatives to end as proof-of-concept projects. Another challenge is that successful patterns and reasons for failure identified in one business operation are not fully reused in other operations, leaving expertise in AI-agent development and operations siloed and resulting in local rather than enterprise-wide optimization. MAAF is a framework for treating AI agents not as standalone automation tools, but as systems that learn and improve alongside business operations and spread the resulting insights to other use cases.

(1) MAS configuration from business knowledge

MAAF ingests not only business manuals and design documents, but also sales discussion and meeting recordings as raw business knowledge. It autonomously identifies what users want to automate and proposes multiple automation options. There is no need to write a formal requirements definition document from scratch. MAAF also provides interactive sessions that, like an interview with an expert consultant, ask only the key questions needed to determine the design. Users can casually share what comes to mind, and the design proposal is refined on the spot. In this way, MAAF produces business-specific MAS whose ability to correctly invoke tools has been verified, enabling multiple AI agents to divide and coordinate complex tasks such as ordering, impact analysis, proposal preparation, and inquiry response.

(2) Safe self-evolution of MAS

Self-evolving multi-AI-agent technology treats MAS construction, operation, and improvement as a single lifecycle. Based on business execution histories and human feedback, it generates improvement candidates for AI-agent prompts, skills, workflows, tools selection, role assignments, and other elements. To prevent "mis-evolution," in which an improvement degrades performance, candidates are verified in an execution environment, and only changes confirmed to be effective are reflected. For important changes, the system incorporates human approval and confirmation and records change histories in an auditable form, balancing continuous improvement with safety.

(3) Continuous expansion of AI utilization

MAAF accumulates successful patterns, reasons for failure, evaluation results, and modification histories obtained through self-evolving multi-AI-agent technology, and organizes them so they can be applied to similar use cases. For example, approaches to exception handling learned from ordering support in retail, impact-analysis procedures learned from system modernization, and proposal-preparation know-how learned from sales support are not confined to their original operations, but are applied to the next AI-agent construction and improvement efforts. This creates a cycle in which experience from the first deployment makes the second more efficient, and experience across multiple deployments improves the quality of AI utilization across the entire enterprise.

We will accelerate the development and operation of AI agents specialized for business operations by linking MAAF with Fujitsu's AI platform, Fujitsu Kozuchi, and Takane, Fujitsu's enterprise generative AI. Going forward, we will apply MAAF to complex and often person-dependent business domains such as ordering operations in retail, investigation, impact analysis, and testing in system development and modernization, and proposal preparation and order booking in sales operations. Through MAAF, we aim to realize a world in which AI agents do not stop evolving after deployment, but serve as partners in business transformation that continue to grow through operation.

Graham Neubig, Associate Professor at Carnegie Mellon University

Overall, the Multi AI Agent Framework seems like a nice method for agent routing and optimization! This is an important topic, as people will be using agents more and more for repetitive tasks, so optimizing the workflows associated with these tasks is a topic that will become more and more top of mind. The approach also seems to be technically appropriate and sophisticated.

Product names and other proper names mentioned herein are trademarks or registered trademarks of their respective owners.