The agents update their own prompts, refine their criteria
In the long arc of human effort to build tools that can keep pace with the world's complexity, Fujitsu has announced a meaningful threshold: AI agents that learn from their own failures, absorb the shifting rules of real business environments, and update themselves without waiting for an expert's hand. Unveiled in May 2026, the technology addresses a quiet but persistent burden — the endless human labor required to keep AI systems current as regulations change, procedures evolve, and institutional knowledge walks out the door. It is, at its core, an attempt to give machines something closer to professional judgment: the capacity to notice what went wrong and decide, on their own, what to do differently.
- Every time a business rule changed or an AI agent failed, a human expert had to diagnose the problem, rewrite the instructions, and push the fix back in — a cycle that never ended and quietly consumed enormous expertise.
- Fujitsu's new multi-agent system breaks that cycle by letting AI watch its own performance, extract lessons from both success and failure, and rewrite its own operating logic without waiting for intervention.
- A built-in self-verification layer prevents the system from drifting into error as it evolves — a critical safeguard in regulated industries where autonomous change without oversight could be catastrophic.
- Fujitsu is partnering with Carnegie Mellon University researchers to make these systems lean enough to run on-premises and at the edge, where sensitive data cannot leave a company's own walls.
- The technology is being positioned as a foundation for what Fujitsu calls sovereign AI — systems that grow smarter wherever they operate, easing workforce shortages and preserving institutional knowledge that retires with its human carriers.
Fujitsu announced this week a new class of AI that can do something its predecessors could not: learn continuously from the work it actually performs. The system deploys multiple AI agents operating as a coordinated team, watching their own results, absorbing human feedback, and adjusting their methods — without waiting for an engineer to intervene.
The problem the technology addresses is deceptively simple to describe and brutally difficult to live with. In any real business, the ground shifts constantly — new regulations arrive, systems get upgraded, procedures change. Teaching an AI to keep pace has historically required something close to constant expert supervision. When an agent failed, humans had to diagnose why, manually rewrite its instructions, and push the corrections back in. The cycle never stopped.
Fujitsu's approach inverts this dynamic. When agents succeed, they extract the reasoning behind that success. When they fail, they analyze what went wrong and propose corrections — then update their own prompts and evaluation criteria accordingly. Crucially, a self-verification mechanism prevents the system from changing itself recklessly, checking its own improvements before deploying them, a safeguard that matters enormously in high-stakes or regulated environments.
The company plans to embed this capability into its Kozuchi AI platform and is collaborating with researchers at Carnegie Mellon University to make the systems efficient enough to run on-premises and on edge devices — environments where confidentiality is paramount and cloud connectivity cannot be assumed. The broader vision, which Fujitsu calls sovereign AI, imagines systems that learn and adapt wherever they operate, easing the burden on specialists, preserving institutional knowledge, and allowing the machine — rather than the person — to carry the weight of constant adjustment.
Fujitsu announced this week a new class of artificial intelligence that does something previous systems could not: learn on its own from the work it actually does. The technology deploys multiple AI agents that operate as a coordinated team, watching their own performance, absorbing feedback from humans, and adjusting their methods without waiting for an engineer to intervene.
The problem Fujitsu set out to solve is deceptively simple to state and brutally difficult to live with. In any real business—a law firm, a manufacturing plant, a financial services company—the ground shifts constantly. New regulations arrive. Systems get upgraded. Procedures change. Documents pile up. A skilled professional knows which information matters, which rules apply, and how to weigh competing priorities. But teaching an AI system to do the same has required something close to constant babysitting. When an AI agent failed at a task, humans had to figure out why. Then they had to manually rewrite the prompts, adjust the search logic, update the evaluation criteria, and push the changes back into the system. This cycle never stopped.
Conventional AI agents, for all their processing power, lacked the ability to diagnose their own failures. They could follow instructions with impressive precision, but they could not look back at what went wrong and decide what to do differently next time. They were, in a sense, locked into their original programming. Adapting them to a changing business environment meant experts had to stay perpetually on call, making adjustments by hand.
Fujitsu's new technology inverts this dynamic. The AI agents now watch themselves work. When they succeed, they extract the reasoning that led to success. When they fail, they analyze the failure, identify what went wrong, and propose corrections. Crucially, they do not just store these insights in a database somewhere. They use them. The agents update their own prompts, refine their evaluation criteria, and incorporate new operational rules—the work that used to demand human expertise and constant attention. The system learns continuously from the actual results of its operations, from feedback humans provide, from policy changes, and from shifts in specifications.
The technology includes a safety mechanism built into this autonomous learning. The agents do not simply change themselves on a whim. They verify their own improvements before deploying them, a form of self-checking that prevents the system from drifting into error. This matters enormously in regulated industries or high-stakes environments where a mistake can be costly.
Fujitsu plans to embed this technology into its proprietary AI platform and eventually into its Kozuchi AI platform, positioning it as a foundational capability for organizations building their own business-specific AI systems. The company is also collaborating with researchers at Carnegie Mellon University—Associate Professor Graham Neubig and Assistant Professor Tim Dettmers—to make these self-evolving systems more efficient, using less memory and power. This efficiency matters because it opens the door to deploying these AI teams not just in cloud environments but in on-premises data centers and edge devices, places where confidentiality is paramount and connectivity cannot be assumed.
The broader ambition is what Fujitsu calls sovereign AI—systems that can learn and adapt wherever they operate, whether in the cloud or behind a company's own firewall. If the technology works as intended, it addresses a cluster of real problems: the shortage of specialists who can keep pace with regulatory change, the loss of institutional knowledge when experienced workers retire, and the difficulty of improving operations when the rules keep shifting. The vision is a workplace where AI and humans learn from each other continuously, where the system grows smarter as conditions change, and where the burden of constant adjustment falls on the machine rather than the person.
Citações Notáveis
Fujitsu aims to realize sovereign AI that can continuously learn not only in the cloud but also in on-premises and edge environments— Fujitsu Limited announcement
A Conversa do Hearth Outra perspectiva sobre a história
What was broken about the old way of doing this?
Every time a business rule changed or an AI agent made a mistake, a human expert had to step in and manually rewrite the system's instructions. It never stopped. The business environment is always shifting—new laws, new procedures, new data—and the AI couldn't adapt on its own.
So the new system teaches itself?
Not quite teaching itself from scratch. It learns from what actually happens when it works. It watches its own successes and failures, figures out what went right or wrong, and updates its own rules. The human is still in the loop, but the burden shifts.
How does it know it's learning the right thing?
It verifies its own improvements before using them. There's a safety check built in. It's not just changing blindly—it's testing the changes against what it knows before deploying them.
Where does this matter most?
Anywhere the rules are complex and constantly changing. Law firms dealing with new regulations. Manufacturing plants with shifting procedures. Financial services with compliance updates. Places where you need someone who understands the nuances but you can't afford to have that person rewriting system instructions every week.
Can it run anywhere?
That's the next phase. Right now it works in the cloud, but Fujitsu is working to make it efficient enough to run on-premises and at the edge—in confidential environments where you can't send data to the cloud. That's where it becomes truly useful for most businesses.
What's the real problem this solves?
The shortage of people who can do this work, and the knowledge that walks out the door when they leave. If the AI can learn and adapt on its own, you need fewer specialists constantly tweaking the system. The knowledge stays embedded in how the AI operates.