Fujitsu has doubled down on self-evolving AI agents by pairing its autonomous multi-agent approach with a new strategic partnership with Anthropic.
For CX executives, the combined message is sharper than either announcement alone. Enterprises have only just normalized 'human-in-the-loop' AI as the safe compromise. Fujitsu is now arguing that the next competitive baseline is AI that improves itself, and it is building the delivery and model partnerships to make that real.
Human-in-the-loop became the enterprise default because it made AI adoption feel controllable. Teams could monitor failures. They could approve changes. They could treat agent improvement like any other change request.
But CX doesn’t wait for quarterly tuning cycles. New policies trigger confusion overnight. Product changes create new intent clusters in hours. A single edge case can become a contact driver by lunchtime.
Fujitsu’s self-evolving multi-agent technology targets that gap. It says agent teams can learn 'continuously and safely' from daily execution results, human feedback, policy revisions, and specification changes. It also claims the system can take over tasks that previously required experts, including prompt adjustments and evaluation criteria updates. In a statement, Fujitsu warned:
“While conventional AI agents demonstrate high processing capabilities for given instructions, they have found it difficult to independently analyze reasons for failure and safely incorporate them into subsequent operations.”
Why the Anthropic Partnership Raises the Stakes for CCaaS and Agent Platforms
The most important new detail is not that Fujitsu picked a frontier model. It is how Fujitsu plans to operationalize it.
Fujitsu says it will strengthen its 'Forward Deployed Engineer (FDE) model' using Claude. The goal is to translate AI into tangible business value through on-site customer collaboration. It also says around 100,000 Fujitsu Group employees will use Claude internally, and it plans to build a 1,000-person engineering team to bring Claude to customers.
This matters for CX because enterprise AI programs typically fail in the messy middle. They stall between prototype and production. They get trapped in governance loops. They struggle to connect to real operational processes and real risk constraints.
Fujitsu is signaling it wants to remove that friction with a delivery engine that embeds into operations. That approach pressures CCaaS and enterprise agent vendors that still depend on a human-managed improvement model. It also raises a new buyer expectation: autonomy is only valuable if a vendor can implement it safely in the customer’s reality. Looking ahead, Yoshinami Takahashi, Chief Operating Officer at Fujitsu argued:
“Through this partnership, we will further strengthen and accelerate our FDE model, ensuring that AI is continuously translated into real value through deep engagement with customer operations.”
Fujitsu Isn’t First to the Concept, but It’s Packaging Autonomy for Enterprises
Self-improving multi-agent patterns have been active across major labs and academia. Stanford has explored multi-agent optimization frameworks that build experience libraries to improve performance. Meta AI has developed collaborative reasoning approaches using synthetic conversations. Amazon has built multi-agent systems for complex enterprise reasoning. OpenAI, IBM, and the LangChain ecosystem have accelerated agent frameworks and evaluation loops.
What has been rare is packaging the full closed-loop improvement concept into an enterprise-ready promise, then pairing it with clear delivery capacity.
Fujitsu is trying to bridge that gap with two moves. It has announced self-evolving agents and also partnered with Anthropic to bring Claude into its stack while emphasizing data sovereignty, regulatory compliance, and security.
Regulated CX Is the Test Case, and the Differentiator Is Trust
Fujitsu’s self-evolving agent release leans into regulated, complex work where rules change constantly and errors have consequences. It also claims an average accuracy improvement of 28 points compared to pre-specialization performance.