Fujitsu Kozuchi Framework Evolves AI Agents With Business Data

Fujitsu began early validation of its Kozuchi Multi AI Agent Framework on July 15, moving beyond experimental AI agent projects and addressing a critical flaw in current deployments. The framework tackles the problem of agents becoming obsolete due to changing regulations or customer needs, often stalling projects after initial proof-of-concept phases. Unlike traditional AI systems, Fujitsu’s system incorporates self-evolving technology designed to continuously improve agents based on results and feedback, expanding insights across the enterprise. This enables companies to treat AI agents as systems that learn and improve alongside business operations, rather than standalone tools.

Fujitsu Kozuchi MAAF: Configuring Multi-Agent Systems from Business Knowledge

Fujitsu is shifting how enterprises use AI with the development of the Kozuchi Multi AI Agent Framework, a system designed to move beyond AI agents towards continuously evolving systems integrated with core business operations. Unlike traditional approaches, Kozuchi incorporates self-evolving technology intended to foster ongoing improvement and expand insights across an entire organization. The core innovation lies in MAAF’s ability to configure multi-agent systems (MASs) directly from existing business knowledge, bypassing the need for extensive, formalized requirements documents.

Fujitsu’s system ingests raw data including business manuals, design documents, and even recordings of sales discussions and meetings, autonomously identifying automation opportunities and proposing options to users. This process is designed to be interactive, mirroring a consultation with an expert; the framework asks key questions to refine designs based on user input, ultimately producing business-specific MAS capable of coordinating complex tasks like order processing, impact analysis, and proposal generation.

A key feature is the framework’s approach to safe self-evolution, treating MAS construction, operation, and improvement as a unified lifecycle. Based on execution histories and human feedback, the system generates potential improvements to AI agent prompts, skills, workflows, and tool selection. To mitigate the risk of detrimental changes, these candidates are rigorously verified in a dedicated execution environment, with only confirmed improvements being implemented.

Importantly, the system incorporates human oversight for critical changes and maintains a complete audit trail, balancing continuous improvement with necessary safeguards. This continuous improvement isn’t limited to individual operations; MAAF accumulates successful patterns, failure analyses, evaluation results, and modification histories, organizing them for application to similar use cases.

Fujitsu anticipates that lessons learned from one area, such as exception handling in retail ordering, can be directly applied to improve AI agent performance in other domains like system modernization or sales support. According to Fujitsu documentation, “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.” The company intends to integrate MAAF with its existing AI platform, Fujitsu Kozuchi, and Takane, its enterprise generative AI offering, accelerating the development and deployment of specialized business agents. Initial applications will focus on complex, person-dependent areas like retail ordering, system investigation, and sales proposal preparation.

Graham Neubig, Associate Professor at Carnegie Mellon University, notes that the framework “seems like a nice method for agent routing and optimization,” adding that “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 important.” Fujitsu aims to establish a future where AI agents are not static tools, but dynamic partners in ongoing business transformation, continuously learning and adapting through operation.

MAAF Self-Evolution: Safe Improvement via Execution History and Feedback

Many artificial intelligence agent deployment projects stall after initial proof-of-concept phases. While enthusiasm for automating business processes with AI agents is high, sustaining those agents through inevitable changes, regulatory revisions, specification changes, or shifting customer needs proves challenging. Existing systems often require substantial redevelopment to remain relevant, effectively treating agents as products rather than adaptive components of a dynamic business environment. Fujitsu addresses this limitation with the newly validated Fujitsu Kozuchi Multi AI Agent Framework, or MAAF, a platform designed to foster continuous improvement and knowledge sharing across an enterprise.

MAAF distinguishes itself through a dual self-evolution process. The framework doesn’t simply automate tasks; it actively learns from execution histories and incorporates human feedback to refine agent performance. This begins with MAS configuration derived directly from business knowledge, ingesting materials like business manuals, recordings of sales discussions, and meeting transcripts.

Fujitsu claims MAAF autonomously identifies automation opportunities and proposes solutions without requiring exhaustive, formal requirements documents. Interactive sessions further refine designs, mirroring expert consultation to quickly pinpoint key elements. The result is business-specific multi-agent systems capable of coordinating complex tasks, from order processing to impact analysis. Crucially, MAAF prioritizes safe self-evolution; only confirmed improvements are implemented, and critical modifications require human oversight, with complete audit trails maintained for transparency and accountability.

This approach balances the benefits of continuous improvement with the need for reliable, predictable performance. The framework’s impact extends beyond individual operations, preventing siloed development and promoting enterprise-wide optimization.