Fujitsu Develops Self-Evolving Multi-AI Agent Technology for Business Operations

Fujitsu has unveiled a groundbreaking self-evolving multi-AI agent technology aimed at revolutionizing how businesses develop and maintain their specialized Large Language Models (LLMs). This new system is engineered to enable a team of AI agents to work cohesively, autonomously handling complex workflows that previously demanded significant human expertise. The technology's primary application is in streamlining the lifecycle of business-specific LLMs. Instead of relying on human experts to meticulously select data, fine-tune learning parameters, evaluate performance, and implement improvements, these multi-AI agents take over. They are designed to continuously learn and adapt by analyzing business execution results and incorporating human feedback, policy revisions, and specification changes. A key feature of this technology is its ability to not just store improvement proposals but to identify the underlying reasons for success or failure, extracting actionable knowledge and operational insights. This allows the agents to generate effective improvement suggestions and verify their impact, ensuring that only beneficial changes are reflected in the model's performance. This continuous optimization process significantly enhances the accuracy of LLM responses and automates tasks like prompt adjustments and evaluation criteria updates, which were traditionally manual and time-consuming. The development signifies a major step towards enabling companies to deploy and refine AI tailored to their unique operations more rapidly and efficiently. By reducing the need for constant intervention from AI specialists, businesses can achieve continuous improvement in their AI systems, ensuring they remain relevant and effective in dynamic operational landscapes. This also extends to tasks like identifying the impact scope of software modifications due to regulatory changes, where AI agents can learn from past cases and human corrections to autonomously improve search and extraction strategies.