NVIDIA and ServiceNow have expanded their enterprise AI collaboration with a focus on autonomous agents that can work across desktops, workflows and data centre infrastructure.
The announcement, made around ServiceNow Knowledge 2026, centres on Project Arc, described by ServiceNow as a long-running autonomous desktop agent for knowledge workers including developers, IT teams and administrators. The important detail for enterprise buyers is not simply that the agent can act across local files, terminals and applications. It is that the companies are positioning governance, auditability and secure runtime controls as core requirements from the start.
That is a useful signal for the wider agentic AI market. As agents move from chat interfaces into real operational environments, adoption will depend on more than model quality. Buyers need to know what an agent can see, which tools it can use, how its actions are contained, how behaviour is monitored and how accountability is preserved when tasks run for longer periods of time.
What was announced
NVIDIA says the expanded collaboration combines accelerated computing, open models, domain-specific agent skills and secure execution software with ServiceNow workflow context, Action Fabric and AI Control Tower governance.
Project Arc is intended to connect natively to the ServiceNow AI Platform through Action Fabric, so autonomous desktop actions can be governed, audited and linked to enterprise workflows. According to NVIDIA, the agent can access local file systems, terminals and applications to complete complex, multi-step tasks, with controls intended for enterprise deployment.
A key part of the stack is NVIDIA OpenShell, an open source secure runtime for developing and deploying autonomous agents in sandboxed, policy-governed environments. ServiceNow is building on and contributing to OpenShell, using it alongside AI Control Tower and Action Fabric to support governed agent execution.
The companies also pointed to open models, agent skills and benchmarking. NVIDIA references its Agent Toolkit, Nemotron open models, the AI-Q Blueprint for specialist research agents and NOWAI-Bench, an open benchmarking suite for enterprise AI agents focused on multi-step workflows.
Why this matters for enterprise buyers
Autonomous desktop agents raise the stakes. A browser-based assistant that drafts text is one kind of risk. An agent that can interact with local files, terminals, enterprise applications and workflow systems is another.
For buyers, the practical questions become sharper:
- Scope: what systems, files, apps and tools can the agent access?
- Permissions: how are actions authorised, limited and reviewed?
- Containment: what prevents a task from reaching data or systems it should not touch?
- Auditability: how are decisions, tool calls and workflow actions recorded?
- Reliability: how are multi-step workflows tested before agents are trusted in production?
- Cost control: how will always-on or long-running agents be measured and optimised?
The ServiceNow and NVIDIA work is notable because it treats these as deployment questions, not afterthoughts. The language around AI Control Tower, OpenShell, Action Fabric and agent observability reflects where enterprise demand is heading: towards agent systems that can be managed like critical infrastructure.
Why suppliers should pay attention
For agent builders, workflow platforms, security vendors and infrastructure providers, the message is clear. Enterprise buyers will increasingly ask for evidence that agents can operate inside controlled environments, not just perform well in a product demo.
That means suppliers need clear answers on runtime security, policy controls, human approval points, observability, logging, identity, data access and operational boundaries. If a product claims to be autonomous, buyers will want to understand exactly how that autonomy is limited, supervised and measured.
There is also an ecosystem angle. ServiceNow and NVIDIA are not presenting autonomy as a single app feature. They are connecting workflow context, secure execution, open models, specialist skills, benchmarking, infrastructure and governance. That points towards a market where the agent stack becomes more modular, and where partnerships between platforms, model providers, security layers and infrastructure companies become commercially important.
The adoption signal
The broader shift is that enterprise AI agents are moving closer to the desktop and closer to the systems where real work happens. That creates more value, but it also exposes more risk.
The next phase of agentic AI adoption will not be won by autonomy alone. It will be shaped by which products can combine useful action with policy, control, audit evidence and enterprise-grade operations. Buyers will want agents that can work across workflows, but they will also want the ability to prove what happened, restrict what is allowed and intervene when necessary.
NVIDIA also frames efficiency as part of the production equation, stating that Blackwell delivers significantly higher token output per watt than Hopper and lower cost per million tokens. Whether buyers run agents on their own infrastructure or through cloud and platform providers, the economic point matters: long-running agents need a cost model that can scale beyond pilots.
The Agentic Expo angle
Agentic Expo is focused on market-ready AI agents. The ServiceNow and NVIDIA announcement shows what market-ready increasingly means for enterprise customers: agents that do useful work, but do it with governance, runtime security, observability, cost awareness and integration into existing workflows.
That is exactly why the agentic AI category needs a dedicated B2B meeting place. Buyers need to compare real deployment architectures, not just product promises. Suppliers need to show how their agents connect, behave, scale and stay accountable. Infrastructure, governance and security are now central to the conversation.
Sources: NVIDIA Blog; ServiceNow press release.