AI
NVIDIA Project Arc on ServiceNow: An Autonomous Desktop Agent You Can Actually Audit
An autonomous agent living on every employee's desktop, completing real work unsupervised, is either the most useful or the most dangerous thing in your estate. The detail that decides which is governance — and that's the whole point of how Project Arc was built.
An autonomous agent living on every employee's desktop, completing real work unsupervised, is either the most useful or the most dangerous thing in your estate. The detail that decides which is governance — and that's the whole point of how Project Arc was built.
ServiceNow and NVIDIA introduced Project Arc at Knowledge 2026 — an enterprise autonomous desktop agent, secured by the NVIDIA OpenShell runtime and governed by ServiceNow AI Control Tower, that lives on employee desktops and autonomously completes complex work. The architecture is deliberately layered: every action runs inside NVIDIA OpenShell, a sandboxed runtime that keeps autonomous activity contained, auditable and enterprise-safe, while AI Control Tower governs the actions — setting policies, monitoring behavior, and logging files read, commands executed and APIs called. ServicenowServicenow
The reason this matters: the agent is powered by ServiceNow Action Fabric and grounded in the CMDB, so its actions are informed by how work actually gets done across workflows, systems and operational history. This isn't a generic desktop bot — it's an agent with enterprise context and an enforced control plane. ServiceNow's framing is that the result is an autonomous desktop agent security leaders can fully audit and approve with confidence. There's also a benchmarking angle worth noting: ServiceNow and NVIDIA released NOWAI-Bench, an open agent-evaluation suite including EnterpriseOps-Gym and EVA-Bench. Servicenow + 2
The honest call. Project Arc is the clearest expression yet of the only safe pattern for desktop autonomy: sandbox the runtime, govern the actions, ground it in real enterprise context. It's also early — treat it as a directional signal of where governed desktop agents are heading, and plan deployment around what's GA and validated, not the keynote. The principle is the takeaway: desktop autonomy is acceptable only when it's contained and auditable by design.
[CTA: Get a readiness review for governed autonomous desktop agents.]
