In most hospitals, someone has the job of making sure there are enough nurses on every floor for every shift, and on a bad morning, working that out can take the better part of a day. Oracle says its software can now do the same work in minutes and, if the hospital allows it, change the schedule without asking anyone first. This is the promise of Fusion Claw, Oracle’s first OpenClaw-inspired runtime that lets AI-powered applications in its Fusion business software handle complex, long-running work within customer-set rules.
What is a runtime? It’s a contained space where a piece of software runs while it does its job. In Fusion Claw’s case, that area is set up for a single job and closed once the work is finished. The AI inside it can reach only the data and tools the customer has allowed it to use.
“It can operate in a deep research mode, where it is doing the modeling, it is doing simulation, could be doing optimization and replanning, and so on,” Kaushal Kurapati, Oracle’s group vice president of product management for Fusion Applications, told The AI Economy in an interview. “But at the same time, it waits for a human being to…review the plan and approve [it]. Or, it can operate also in a fully autonomous mode if it’s been given that permission, given the authority from the human being.”
Fusion Claw isn’t designed for individual tasks, but rather for delivering business outcomes. “What we realized…last year is enterprise work is very predetermined,” he said, meaning that most processes follow a known recipe and steps are set ahead of time. Agents are instructed to carry them out the same way every time. But now with larger enterprise apps, the focus is more on problems where the goal is clear, but the steps aren’t—getting hospital staffing to a certain level, for example. “That is the business outcome which enterprise apps are meant to solve in the agentic era,” Kurapati added. “Within that, there is a piece of work…which involves a lot of taking into constraints, doing deep research, simulation, optimization, [and] long-running workflows. For that, we’ve specifically introduced Fusion Claw.”
Like OpenClaw, But Built From Scratch
And to be clear: Although it’s described as being OpenClaw-inspired, Kurapati emphasized Fusion Claw was built from scratch, with Oracle’s architecture and proprietary IP. Still, its launch now puts the company in a growing list of enterprise tech vendors offering agentic solutions built on or inspired by OpenClaw. Nvidia, Microsoft, Automation Anywhere, and Salesforce have all moved to bring OpenClaw-style agents into the enterprise, though none appears to have built anything like it.
Fusion Claw draws on what’s become known as the “Claw” class of AI agents, named after OpenClaw, the open-source project that went viral earlier this year. OpenClaw is known for taking action on its owner’s behalf, and with permission, it can answer emails, manage calendars, work with files, and carry out tasks while the person is away. That same reach has drawn security concerns, because an agent with broad access to someone’s accounts can do real damage if it’s misconfigured or tricked. Other companies have since built on OpenClaw to address those worries, including Nvidia, whose NemoClaw wraps the agent in a secure runtime called OpenShell. So why did Oracle elect to build its own instead of adopting an existing Claw?
Kurapati said that when Oracle looked into the different Claws, it evaluated two areas: security, safety, and trust, and “the ability to execute enterprise workflows deterministically.” “We don’t believe that other implementations were particularly suited for and measured up to our criteria when we looked at it,” he explained. “We wanted to be sure that this container and Claw that executes an agentic runtime, that executes with our agentic apps, is highly secure in the sense that it comes into being only when the task is given. It orchestrates the work, plans, learns, and goes away when it doesn’t need to exist anymore once the task is complete.”
How Fusion Claw Works
Customers don’t use Fusion Claw directly. They work through Claw-powered applications, specialized programs built around a particular business goal. Those apps call on the runtime whenever a job involves heavy research, simulation, or planning. Oracle has released 25 of them so far, covering HR, finance, supply chain, and sales. Workforce Staffing is the kind of app behind the hospital scenario presented earlier, while Ledger helps finance teams reconcile accounts and investigate discrepancies. Shipping Consolidation helps teams discover ways to reduce shipping costs without missing delivery commitments. Lastly, Account Territory Growth Plan helps sales leaders decide how to divide up territories. Kurapati said more are on the way, both from Oracle and from customers and system integrators building their own.
To keep Fusion Claw within bounds, customers define what Oracle calls an Enterprise Operating Envelope, essentially the rulebook for how an organization operates. It covers the company’s goals, standard operating procedures, and policies, along with who can make which decisions, how much risk is acceptable, what requires a manager’s approval, and when a problem should be handed to a person. Some of those rules can come from documents the company already follows. In Oracle’s hospital staffing scenario, the administrator provided Fusion Claw with a staffing procedures document that included a rule stating that any plan leaving coverage below 95 percent could not be applied automatically. Fusion Claw uses an AI model to translate documents like that into code, so it applies the rules the same way every run.
The envelope sets the rules, and Oracle’s Outcome Trust Harness makes sure Fusion Claw follows them. Each time Fusion Claw takes on a job, the harness limits it in four ways: whose authority it acts under, what it’s allowed to do, what data it can see, and what actions it can take. In practice, Kurapati said, each job runs in a private, isolated workspace that only the app requesting the work can open. That workspace has the same access permissions as the app’s agent and can’t access the internet on its own. If a company wants Fusion Claw to use an outside tool, it must connect that tool to the agent first. Once the job is done, the workspace shuts down.
Once the work is finished, Fusion Claw generates an Outcome Receipt, a record of what happened that organizations can show to auditors or a governance committee. It details the authority that was applied, the evidence used, the decisions made, the actions and transactions carried out, and the final result. In the hospital scenario, the receipt would show who requested a new schedule, explain why Fusion Claw chose that particular course of action, and point to the staffing document it followed. It might also specify that no approver was needed because Fusion Claw acted within the authority it had been given.
For Kurapati, those safeguards are the point. When companies evaluate an agent that can act on their behalf, he said, what they want is predictability. “They want things executed deterministically. They want things executed at scale. They want things executed in the policy framework of their organization,” Kurapati said. “They don’t care…whatever else it has learned on the internet. They want to make sure that it follows the procedural document [they] gave it.”
A Different Way to Pay for Enterprise AI
Safeguards like these address whether a company can trust the software. The next question is what it will cost to run. Kurapati highlighted that Fusion Claw is billed through Oracle’s consumption-based pricing, which uses what the company calls AI units, a measurement of business value and the business outcome customers ask an agent to achieve. The AI units model is not new—it’s part of the platform’s Fusion offering.
“If you’re doing a simple summarization, you’ll use fewer AI units, and if you’re asking it to execute a full task autonomously, it’ll use a higher amount of AI units because it’s a much higher value task,” Kurapati explained. “We measure based on the business outcome that an agent is trying to accomplish, and the units are appropriately valued.”
Oracle said customers will consume AI units when an LLM is involved in planning. Once a plan is developed, Oracle’s Fusion platform executes it. At this point, no AI units are consumed. “Whether it’s processing massive data sets, computation, calculations, all of the simulations, [and] things like that, you don’t need an LLM anymore,” Kurapati said. If the results are what’s expected, the job is done. If the AI didn’t meet the requested conditions, it cycles back to learn, adapt, and change the conditions to produce the desired outcome. But remember: When you’re involving an LLM, AI units are spent.
He added that the AI learns from each job, recording the path it took and noting what worked and what didn’t. The next time a customer runs the same or a similar problem, it can follow what worked before instead of reasoning through everything from scratch. Because AI units are consumed while the model does that reasoning, Kurapati said the system “becomes more efficient over time.”
Fusion Claw runs on Oracle Cloud Infrastructure and is powered by leading AI models, including those from Google and OpenAI. Kurapati said that it has already started rolling out to early-access customers and will become more broadly available to everyone else starting the first week of October.
Discover more from Ken Yeung
Subscribe to get the latest posts sent to your email.
