Oracle Tuesday introduced Fusion Claw, a runtime that lets its Fusion cloud applications carry out multi-step business tasks on their own, and released 25 applications built on it, SiliconANGLE reported. The launch brings Oracle's catalog of Fusion Agentic Applications to 75.

What Happened

Claw pairs large language models, which handle analysis and planning, with conventional software that performs calculations and processes transactions. Heavy number crunching over large data sets happens outside the model, which Oracle says reduces reliance on expensive model calls. The runtime is built for long-running work that involves modeling, optimization and iterative planning.

It connects to outside systems through APIs, the Model Context Protocol and the Agent2Agent protocol. The model runs in an isolated environment and cannot directly modify Fusion business objects, limiting the risk of unauthorized changes to records.

RelatedGoogle Cloud Unveils Gemini Agent That Gets Its Own Work Identity Inside Companies

When a task finishes, Claw generates an "Outcome Receipt" that shows the policies applied, including their full text, the decisions made and the transactions performed. Customers define an Enterprise Operating Envelope that sets objectives, permissions, risk thresholds and approval requirements. Agents can either propose plans for an employee to review or act automatically within delegated authority. CIO.com reported that a second control, the Enterprise Trust Harness, maintains governance boundaries and produces auditable records.

At launch Claw supports Google Gemini and OpenAI models. Oracle plans to add others but has not said which, and customers cannot yet bring their own open-source models.

The 25 new applications cover jobs such as investigating accounting exceptions, building workforce staffing plans, consolidating shipments and planning sales territories. In the staffing case, an agent weighs qualifications, availability, labor rules and cost to propose schedules, and reopens the plan on its own when someone calls in sick. SiliconANGLE said all 25 are available now.

Why It Matters

Oracle is betting that enterprises will trust agents in finance, HR and supply chain systems only if the model is kept at arm's length from the books. Routing calculations and postings through deterministic code addresses two common objections to generative AI in ERP: that models make arithmetic errors, and that per-step inference makes costs hard to predict.

Natalia Rachelson, Oracle's senior vice president of Fusion Applications, told SiliconANGLE that "any task can be in full auto mode," though she expects customers to start with more human review and widen automation over time.

Pricing and the Cost Question

Claw is included at no extra charge with Oracle's agentic applications product, which is sold separately from standard Fusion. Customers pay for the AI units their work consumes.

Analysts quoted by CIO.com were cautious. Scott Bickley of Info-Tech Research Group said separating reasoning from execution improves reliability, but noted that total cost still depends on the number of actions, token use, model choice, testing, human review and support. Manoj Chandra Jha of Nord-IQ said governance work grows even as development gets simpler, and advised buyers to get per-outcome pricing in writing and run bounded pilots.

What to Watch

Early customers will show how often agents run in automatic mode versus review mode, and what AI unit bills look like for tasks like territory planning or staffing. Oracle's choice of additional supported models, and whether it opens Claw to customer-supplied models, is another open item as rivals such as SAP and Workday build their own agent layers.