ServiceNow makes Arc generally available, a work agent policed by NVIDIA OpenShell

What ServiceNow launched

ServiceNow has made Arc, a personal AI agent for office work, generally available. The company announced the launch on 21 October 2026 at NVIDIA's GTC conference in Berlin, in a press release that describes Arc as "the personal long-running AI agent built for the enterprise." ServiceNow and NVIDIA first outlined the agent in May, according to ZDNet as cited by AI Affairs; Wednesday's news is that customers can now switch it on.

The pitch is less about what Arc can do than about how tightly it can be fenced in. Rules for the agent are set in ServiceNow's governance product, which it calls the Super Intelligence Control Tower, and enforced while the agent runs by NVIDIA OpenShell, the open-source agent runtime NVIDIA introduced this month. On the same day NVIDIA announced an OpenShell Reference Architecture that pairs that software with its Vera Rubin servers, Vera CPUs and BlueField data processing units, and it named ServiceNow's Arc as a product built on it.

That makes Arc one of the first commercial agents whose maker leads with the containment story rather than with benchmark scores. Whether that is enough to persuade IT departments to let an autonomous agent loose in their inboxes and ticket queues is the real test.

What Arc does

ServiceNow describes Arc as long-running: instead of answering one prompt, it takes a goal, plans how to reach it, works across enterprise systems, checks its own output and keeps going until the task is finished, "whether that's minutes or days." It works across ServiceNow workflows, email, Microsoft Teams and other common productivity tools.

The company lists five headline abilities. Arc is meant to own a task end to end once given a goal. It surfaces work proactively, pulling assigned tasks, pending approvals and needed follow-ups out of an employee's backlog, inbox, meetings and conversations. It can build an integration on the fly when one does not exist, which ServiceNow says it runs inside the same OpenShell boundary so it cannot reach further than the organisation allows. It builds memory and skills over time, so performance should improve the longer it works with one person. And it can split work between local and cloud models.

That last point ties Arc to NVIDIA's new laptops. ServiceNow says Arc is designed to handle everyday work with local models on NVIDIA RTX Spark machines and to call frontier models only for more demanding tasks, to cut inference costs and keep more data on the device. RTX Spark laptops went on sale last week, so the local option depends on companies buying new hardware. Arc can also run on its own computer inside a secure OpenShell runtime instead of on an employee's desktop.

How the controls are meant to work

The governance design has two layers. In the Control Tower, IT staff define policies: which systems an agent may reach, which APIs it may call and which data it may read. OpenShell then reads those policies and applies them while the agent is running, so the limits hold at the moment of action rather than only in a configuration screen.

ServiceNow says the controls can be fine-grained. An organisation could, for example, block a single action such as writing to a production system while letting the agent carry on with the rest of its work. That matters for an agent that strings together dozens of steps, because an all-or-nothing switch would make it useless the first time a risky step came up.

OpenShell itself is a sandbox: a runtime boundary an agent works inside, with what it can reach set by policy. We covered NVIDIA's original OpenShell launch earlier this month, and on Wednesday Red Hat said it would bring OpenShell and BlueField monitoring to OpenShift. ServiceNow's contribution is the policy source: the rules come from a system many large companies already use to manage IT requests and approvals.

"As AI agents take on more autonomous tasks, governance and security have to extend to everything they touch," said Jon Sigler, ServiceNow's executive vice president for its AI Platform. NVIDIA's Justin Boitano said OpenShell lets ServiceNow extend its governance "down into the runtime itself, the layer where agents actually take action."

AI Passport and the hardware kill switch

The most ambitious part of the announcement is not shipping yet. ServiceNow's AI Passport, announced the week before, is what the company calls an attested record issued to each AI agent, MCP server and other AI asset the Control Tower governs. It carries business rules such as risk tolerance and which outside services an agent may use, so that enforcement points can check what an agent is trusted to do.

Under the plan, NVIDIA OpenShell will read those passports in software, and NVIDIA Sentry running on BlueField data processing units will read them in hardware. If IT revokes an agent's passport, ServiceNow says BlueField can enforce that "at machine speed" before the agent proceeds, with the revocation spreading across both layers "within approximately one second." Only the revoked agent's traffic would stop; other work on the same server would continue, and restoring it would need no restart. The aim is a stop button that compromised software cannot override.

The catch is in the availability section of the release: AI Passport, including enforcement in OpenShell and BlueField, "is expected to be available in a future release." Both TSN and AI Affairs point out that the one-second figure describes a planned system, not something included with Arc at launch. For now, Arc's guardrails are the Control Tower policies enforced by OpenShell in software.

Where Arc fits among personal agents

Arc lands in a crowded week for personal agents. Anthropic opened Claude Duet, an app that watches connected tools and asks before acting, on the same day. OpenAI has Dots, its always-on agents, and Meta is pushing Muse to consumers. All of them promise an assistant that notices work and does it without being asked each time.

ServiceNow's angle is that it already sits where much corporate work is approved and tracked. An agent that can see a pending approval in the ServiceNow queue, a related email thread and a Teams message, and is governed by the same system that logs IT changes, is easier to justify to a security team than a consumer app with broad account access. The flip side is that Arc is tied to ServiceNow's platform and pricing, while the consumer agents are not.

NVIDIA, for its part, gets a flagship customer for OpenShell. Its reference-architecture release lists a long roster of partners, from Anthropic and OpenAI to CrowdStrike and SAP, but much of the wording is "working with" or "exploring". TSN notes that ServiceNow is the clearest case of a shipping product.

Price, platforms and who can use it

Arc is part of ServiceNow EmployeeWorks and is included at no additional cost with the Foundation, Advanced and Prime versions, according to the release. That means it is not a standalone free product: it comes with those existing subscriptions, whose prices ServiceNow does not list publicly.

Mac support is live now, and ServiceNow says Windows support is expected to roll out within weeks. The release does not say whether Arc is available in every region on day one, or which local and frontier models it uses by default, beyond the general split between local inference on RTX Spark machines and frontier models for harder tasks.

For IT teams, the practical first steps are setting Control Tower policies before switching Arc on, deciding which systems it may write to, and choosing whether it runs on employees' desktops or on its own machine inside OpenShell. Both deployment modes are described in the release; neither is presented as the default.

What is still unproven

ServiceNow calls the combination "industry-first" and Arc "the first personal AI agent built for everyday work and governable by IT." Those are the company's own claims, and similar governance features are arriving from several vendors at once.

There are no customer results in the announcement: no figures for time saved, tasks completed or error rates, and no description of how Arc behaves when a policy blocks a step midway through a multi-day task. The ability to build its own integrations is powerful, but it is also exactly the kind of behaviour security teams worry about; ServiceNow's answer is that anything Arc builds runs inside the same OpenShell boundary, which customers will want to test.

Agents that can act for days also raise questions that policies alone do not settle, such as who is accountable when an agent sends the wrong email or approves the wrong request. Arc's controls narrow what can go wrong. The hardware-backed revocation that would make them harder to bypass is still on the roadmap, and its one-second promise will need to be shown in real deployments.