Most AI agent projects fail after the initial proof-of-concept phase because they can't adapt when business needs change. Fujitsu's newly validated Kozuchi Multi AI Agent Framework (MAAF) tackles this problem head-on by creating AI agents that learn and evolve on their own, rather than becoming obsolete when regulations shift or customer demands evolve.

Why Do Most AI Agent Projects Stall?

The challenge facing enterprises today is straightforward but stubborn: AI agents are typically built as static tools. When business requirements change, regulations update, or customer needs shift, these systems require substantial redevelopment to stay relevant. Many companies treat their AI agents like finished products rather than living systems that should adapt alongside the business itself. This disconnect between static AI systems and dynamic business environments is why so many projects stall after initial enthusiasm fades.

Fujitsu's approach flips this model entirely. Instead of requiring teams to rebuild agents from scratch when conditions change, MAAF incorporates what the company calls "self-evolving technology" that continuously improves agents based on execution history and human feedback. The framework learns from what works and what doesn't, then proposes refinements to agent prompts, skills, workflows, and tool selection.

How Does Fujitsu's Self-Evolving System Actually Work?

The framework operates through a carefully balanced process that prioritizes safety alongside continuous improvement. When potential improvements are identified, they're rigorously tested in a dedicated execution environment before being deployed to live systems. Critical changes require human oversight, and the entire system maintains a complete audit trail for transparency and accountability. This approach ensures that agents improve without introducing unpredictable failures.

What makes MAAF particularly innovative is how it ingests business knowledge directly. Rather than requiring exhaustive, formal requirements documents, the system can work with raw materials like business manuals, design documents, sales meeting recordings, and transcripts. It autonomously identifies automation opportunities and proposes solutions in an interactive consultation-style process that mirrors working with an expert advisor.

Steps to Implementing Continuously Evolving AI Agents

• Knowledge Ingestion: Feed the system existing business materials including manuals, design documents, meeting recordings, and sales discussions rather than starting from formal requirements documents.
• Interactive Refinement: Work through a consultation-style process where the framework asks clarifying questions to understand your specific business needs and constraints before designing multi-agent systems.
• Safe Testing and Deployment: Allow proposed improvements to be verified in a dedicated execution environment before implementation, with human oversight required for critical changes and complete audit trails maintained.
• Cross-Domain Learning: Enable the system to accumulate successful patterns, failure analyses, and modification histories so lessons from one business area can improve performance in other domains.

Fujitsu began early validation of MAAF on July 15, moving beyond experimental projects to address real-world deployment challenges. The framework is designed to handle complex, person-dependent business processes like retail ordering, system investigation, and sales proposal preparation. Initial applications will focus on these high-value areas where human expertise has traditionally been difficult to automate.

One of the framework's most powerful features is its ability to share learning across the entire enterprise. Rather than treating each AI agent deployment as an isolated project, MAAF accumulates successful patterns, failure analyses, evaluation results, and modification histories. This means experience from automating exception handling in retail ordering can directly improve AI agent performance in system modernization or sales support.

"This 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 important," noted Graham Neubig, Associate Professor at Carnegie Mellon University.

Graham Neubig, Associate Professor at Carnegie Mellon University

Fujitsu plans to integrate MAAF with its existing AI platform, Kozuchi, and Takane, its enterprise generative AI offering. This integration will accelerate how quickly organizations can develop and deploy specialized business agents tailored to their specific needs.

What Does This Mean for the Future of Enterprise AI?

The broader implication is a fundamental shift in how enterprises should think about AI deployment. Rather than viewing AI agents as one-time implementations, organizations can now treat them as dynamic partners in ongoing business transformation. These systems continuously learn and adapt through operation, improving over time rather than degrading as business conditions change.

This development arrives as the U.S. government is also prioritizing quantum computing and AI advancement. The Department of Defense recently approved Vibrint's Quantum Circuit Factory and Secure Light Fidelity solutions through its Tradewinds Solutions Marketplace, making advanced quantum development tools more accessible to federal agencies. This broader push toward democratizing advanced technologies suggests that frameworks like Fujitsu's MAAF represent a larger trend toward making sophisticated AI systems more practical and deployable across organizations.

The key takeaway is that AI agents no longer need to be treated as static tools that become obsolete when business needs change. With self-evolving frameworks like MAAF, organizations can build AI systems that improve alongside their operations, turning AI from a one-time implementation into a continuous source of competitive advantage.