ServiceNow and NVIDIA Launch Project Arc Desktop AI Agent

ServiceNow and NVIDIA used Knowledge 2026 to push agentic AI past the chatbot. The two companies unveiled Project Arc, a new autonomous desktop agent that thinks, writes code, executes actions across enterprise tools, and adapts when something breaks — all without a pre-built workflow telling it what to do.

The catch, and the reason this matters, is where it runs: inside NVIDIA OpenShell, a sandboxed runtime that boxes the agent in with policy controls, full audit logging, and real-time governance from ServiceNow's AI Control Tower.

What Is Project Arc

Project Arc is a desktop AI agent that can complete complex, multi-step work across enterprise tools — opening apps, writing code, calling APIs, extracting and reformatting data, and stitching together actions that would normally require a human to switch between five systems. Unlike scripted automation, it improvises a path through tasks rather than following a fixed script, and it adjusts when something doesn't go as planned.

That's the same pitch every "computer use" agent has made over the past year. What's different here is the runtime.

Every action the agent takes runs inside NVIDIA OpenShell, a sandboxed environment that adds policy-based management so autonomous activity stays contained, auditable, and enterprise safe. ServiceNow AI Control Tower governs the actions the agent takes, sets policies, monitors behavior, and logs files read, commands executed, and APIs called. If the agent does something it isn't supposed to do, the runtime catches it before it touches a production system.

Project Arc connects natively to the ServiceNow AI Platform through ServiceNow Action Fabric, an open system that lets any AI agent — including Arc — execute governed work on the platform.

Why This Matters

Desktop agents have been the obvious next leap in enterprise AI for the better part of two years, and most pilots have run aground on the same wall: governance. CIOs love the idea of an agent that can fill out forms, reconcile records, and run end-to-end processes across tools. They are far less excited about a Python-driven assistant clicking through their financial reporting system without a paper trail.

OpenShell is the answer to that wall. By isolating the agent in a sandbox, logging every action, and giving the security team a kill switch, ServiceNow and NVIDIA are trying to make autonomous desktop work auditable in the same way credit card transactions are auditable.

That changes the conversation inside large enterprises. Rather than ask "should we let an AI do this work," the question becomes "what controls do we want around this agent." That's the conversation security and risk teams already know how to have.

How OpenShell Works

OpenShell is NVIDIA's sandboxed runtime for autonomous AI. It runs the agent's actions inside a controlled environment, with policy enforcement at the boundary. Files the agent reads, commands it executes, and APIs it calls are all logged. ServiceNow's AI Control Tower sits on top of OpenShell, defining what the agent can and cannot do based on the user's role, the system being touched, and the workflow it's running.

When an action falls outside policy — say, the agent tries to write to a system the user is not authorized to modify — Control Tower can block the call in real time. Critically, it can also shut the agent down entirely if behavior drifts, the runtime equivalent of a circuit breaker.

The architecture decouples the agent's reasoning model from the runtime that executes its actions. That separation is what makes the system enterprise-ready. The reasoning layer can change as models improve. The governance layer stays the same.

Industry Impact

Project Arc lands in a crowded field. OpenAI has Operator. Anthropic ships computer use as a Claude capability. Google has its own desktop agent on Gemini. Microsoft, Perplexity, and a growing number of startups are pitching variants. What ServiceNow and NVIDIA are betting is that enterprise customers don't want yet another standalone agent — they want one that plugs into the system of record they already have, with the governance their auditors already trust.

That bet aligns with the day's other ServiceNow news. The company also expanded its AI Control Tower to govern agents running on AWS, Google Cloud, and Microsoft Azure — and to govern Microsoft Agent 365 agents specifically. The unified message: ServiceNow wants to be the agent control plane regardless of where agents come from or where they run.

For NVIDIA, OpenShell is a quiet but strategic move. Most attention on NVIDIA in the enterprise has focused on chips and the AI Factory. OpenShell extends the company's reach into the runtime layer, where policy and security live. It's a place customers don't typically expect to see NVIDIA, and that's part of the point.

What People Are Saying

Fortune covered the announcement under the framing "an AI workforce that can run your entire company," quoting ServiceNow CEO Bill McDermott on the need for AI that "senses, decides, and securely acts." Industry analysts attending Knowledge 2026 highlighted the contrast with consumer-style agent demos — the focus here is squarely on what large enterprises require to deploy agents at scale.

Some observers raised the obvious concern: a sandbox is only as good as its policies. If Control Tower is configured loosely, a permissive policy can still let an agent reach production data it shouldn't touch. That puts the burden back on the customer's security team to define guardrails carefully — not a new problem, but one Project Arc inherits rather than solves.

What's Next

Project Arc is available as an early preview. The AI Control Tower integration with NVIDIA's Enterprise AI Factory validated design is generally available now. ServiceNow expects pilots across customer service, IT operations, and finance — the same enterprise functions that are already running its agents today, now extended onto the desktop.

If Project Arc holds up under real workloads, expect competitors to follow with their own sandboxed runtime stories. The desktop agent race is moving from "can it click the button" to "can your audit team approve that it clicked the button." That's the harder question, and the one that decides which agents make it past the pilot.