Fujitsu tests agents that improve themselves from real business feedback

What changed: Fujitsu announced a self-evolving multi-agent technology that lets teams of AI agents learn from execution results, human feedback, policy changes, and specification updates instead of waiting for specialists to rewrite prompts and rules. Fujitsu says it applied the system to business-specific LLM improvement across manufacturing, healthcare, finance, and public administration, reporting a 28-point average accuracy improvement versus pre-specialization performance.

Why it matters: For builders, the useful idea is not “fully autonomous AI,” but a safer loop for keeping agents current when business rules change. If this works in production, it points to fewer brittle automations and less ongoing expert maintenance.

Try/watch: Watch whether Fujitsu exposes this as a product teams can buy, not just a research capability, and ask how failed agent changes are reviewed before they affect live workflows.

Blue Yonder adds supply-chain agents that explain ordering decisions

What changed: Blue Yonder announced new cognitive supply-chain products plus new and expanded agents, including Inventory Ops Agent skills, an agentic ordering workflow for supplier order approvals, and a Workforce Management Knowledge Agent for configured-solution support. The update also adds clearer data-source citations, auditability, scenario modeling, and planning features intended to show users why a recommendation was made.

Why it matters: Operators do not need another chatbot; they need agents that can help approve orders, explain tradeoffs, and route people toward better decisions during shortages, warehouse changes, or demand swings. This is a practical example of agentic AI moving into industry software where explainability matters as much as automation.

Try/watch: If you run supply chain, retail, or manufacturing ops, test agents first on recommendation-and-review workflows before letting them auto-approve supplier or inventory actions.

Agent security advice shifts from “better prompts” to system-level controls

What changed: CSO Online covered new research arguing that enterprises should treat the AI model inside an agent as untrusted and enforce safety around the whole system, especially once agents can use enterprise tools, memory, browsers, and business applications. The article highlights five security principles for agents: least privilege, tamper resistance, complete mediation, secure information flow, and treating the human as a weak link.

Why it matters: This is a useful correction for buyers and consultants: prompt filters alone are not enough when an agent can read files, move data, send messages, or trigger workflows. Agent projects should be scoped like access-control projects, with logging, isolation, approvals, and limits on what each agent can touch.

Try/watch: Before expanding an agent pilot, write down what data the agent can access, which actions require human approval, and how you would reconstruct what happened after a bad action.

Informatica and Databricks push governed data into agent workflows

What changed: Informatica announced new Databricks-related capabilities, including headless data management through Model Context Protocol servers, which let agents invoke Informatica services such as metadata search and address validation inside workflows without custom integration work. The same announcement includes Lakebase connectivity designed for agentic use cases, master-data publishing into Databricks, and Unity Catalog tag extraction for governance.

Why it matters: Agents are only as useful as the data they can safely reach. For enterprises, this points to a more realistic path: connect agents to trusted records, metadata, and governance tags instead of asking them to reason over messy exports and undocumented tables.

Try/watch: If you are building internal agents, prioritize clean access to customer, supplier, product, and policy data before adding more models or front-end chat experiences.