Oracle expands AI agent features with Fusion Claw
SiliconANGLE Paul Gillin
Oracle just gave Fusion apps a new AI runtime for bigger business jobs, from payroll planning to accounting cleanup. It’s built to keep the AI out of the money-moving parts, which is the whole point.
Based on reporting by SiliconANGLE, Paul Gillin — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Oracle is widening the job description for its Fusion agents. The company says its new agentic software runtime, Fusion Claw, can handle longer-running business work inside Fusion applications, including accounting reconciliation, staffing plans, shipment consolidation and sales territory planning.
This is a step up from the earlier agent tools Oracle shipped for narrower goals. Those were built for things like renewals. Claw is meant for messier work that changes as it goes, where the system has to model options, optimize them and then revisit the plan when reality shifts. Oracle’s staffing example is a good one: the agent can weigh qualifications, availability, labor rules and cost targets, then propose schedules and, if allowed, update the system of record. If someone calls in sick, it can reopen the plan.
The company is also trying to keep a lid on the cost of all this. Oracle says Claw uses a language model only when it needs reasoning, while calculations and transactions over large data sets happen outside the model. The first release supports Google’s Gemini and OpenAI models, and Oracle says more will come later, though it isn’t saying which ones. Users still can’t bring their own open lower-tier models for Claw tasks.
Oracle says the new agent layer is available in 25 applications right now, bringing the total number of Fusion Agentic Applications to 75. The new tools can also connect to third-party systems through APIs, Model Control Protocol and Agent2Agent. But the real selling point is control: customers set objectives, permissions, risk thresholds and approval rules inside what Oracle calls an Enterprise Operating Envelope, then choose whether a person reviews the plan or the app executes within delegated authority.
And Oracle is clearly trying to make that handoff feel less scary. Claw runs in an isolated environment, so the language model can’t directly change Fusion business objects. When a task ends, Oracle generates an Outcome Receipt that shows the policies used, the decisions made and the transactions performed, with more context than a normal audit log. The company says customers will likely start with human review and loosen the reins over time. Oracle did not share measured cost savings or production evidence, which leaves the pitch sounding promising but still a little theory-heavy.
My take — AI-written commentary, not fact-checked reporting
This is the right direction: keep the model away from the ledger, the roster and the receipts, then let it help plan the mess. The industry keeps pretending “agentic” means “trust it with the keys”; Oracle is at least admitting that most companies want a chaperone first. That’s not a buzzword, that’s basic self-preservation.
Read more about this at: SiliconANGLE
