Fujitsu announced on July 13 that it will begin early verification of a platform called Kozuchi Multi AI Agent Framework, or MAAF, starting July 15. The framework auto-generates groups of AI agents from business knowledge and continuously improves them based on execution results and human feedback, according to IBTimes JP.
The platform targets a persistent bottleneck in enterprise agent deployments: business procedures that combine exceptions, judgment criteria, and system operations are difficult to capture in a single build. Fujitsu said that regulatory changes, specification updates, and shifting customer requirements have kept many enterprise AI projects stuck at proof-of-concept stages.
How MAAF Works
MAAF ingests business manuals, design documents, and recordings of commercial discussions and meetings as operational knowledge. From that input, it proposes multiple automation plans by identifying which workflows should be automated. An interactive session function asks focused follow-up questions on key design issues, reducing the need for formal requirements documents before agent construction begins.
The framework then produces multi-agent systems verified for correct tool invocation. In operation, multiple agents divide roles and coordinate on tasks including ordering, impact analysis, proposal preparation, and inquiry handling.
Self-Evolving Agent Lifecycle
The core technical differentiator is what Fujitsu calls “self-evolving multi-AI agent technology.” Construction, operation, and improvement of a multi-agent system are treated as a single lifecycle. Based on execution histories and human feedback, the system generates candidate improvements for prompts, skills, workflows, tools, and role assignments, according to IBTimes JP.
Candidate changes are tested in an execution environment, with only validated modifications reflected in the live system. For high-impact changes, human approval is required, and change histories are retained in auditable form.
Initial Applications and Integration
Fujitsu plans to link MAAF with its Kozuchi AI platform and Takane, its enterprise generative AI offering. Initial trial applications include retail ordering operations, investigation and impact analysis in system development and modernization, and proposal preparation and sales booking.
Graham Neubig, an associate professor at Carnegie Mellon University, described the framework as “a strong approach to routing and optimization for AI agents,” according to IBTimes JP.
The Auto-Building Market Shift
The MAAF trial arrives as enterprises increasingly confront the gap between agent capability and agent deployment complexity. Gartner found that in 2024, 60% of GenAI proofs-of-concept were abandoned upon completion, a figure Gartner projects will fall to 35% by 2029. Auto-forming frameworks like MAAF represent one approach to closing that gap: instead of hand-coding agent systems from scratch, organizations can feed in their existing documentation and let the platform generate, test, and iterate agent configurations.
Whether automated agent construction can reliably handle the edge cases and judgment calls that make enterprise workflows difficult in the first place is the open question Fujitsu’s trial will need to answer.