NVIDIA and ServiceNow team up on desktop AI agents

AI can write your emails and debug small snippets of code, but letting an autonomous model take over your actual desktop terminal is usually enough to give enterprise IT directors a panic attack.

At ServiceNow Knowledge 2026, NVIDIA CEO Jensen Huang and ServiceNow CEO Bill McDermott took the stage together to change that mindset. The two companies unveiled Project Arc, an autonomous, self-evolving desktop agent designed for knowledge workers like developers, IT staff, and system admins.

Sandboxing the desktop agent

Project Arc is built to handle long-running, multi-step tasks right on a user's machine. Unlike standalone chat tools, it can interact directly with local file systems, open terminals, and operate installed enterprise software.

To keep an autonomous worker from going rogue, the agent connects natively to the ServiceNow AI Platform through ServiceNow Action Fabric. This adds workflow context, audit logs, and compliance tracking to every single step the desktop agent takes.

Under the hood, Project Arc runs on NVIDIA OpenShell, a newly open-sourced secure runtime designed for sandboxed agent execution. OpenShell essentially gives IT teams a strict policy engine to dictate exactly what files an agent can read, what local tools it can touch, and how tightly contained its actions remain. Management happens through ServiceNow AI Control Tower, giving enterprises real-time observability across the agent lifecycle.

Benchmarks and the cost of tokens

Deploying autonomous agents around the clock also brings massive computational challenges. If an agent runs constantly in the background, token bills can spiral quickly.

NVIDIA emphasized that running Project Arc on its Blackwell platform yields over 50 times greater token output per watt compared to Hopper. According to NVIDIA, that performance leap cuts the cost per million tokens by nearly 35 times in production AI environments.

On the software side, the partnership is pushing NOWAI-Bench, an open benchmarking suite tailored to multi-step enterprise workflows using the NeMo Gym library. On its EnterpriseOps-Gym benchmark, NVIDIA Nemotron 3 Super currently ranks as the top open-source model.

Why it matters

We think this partnership tackles the exact roadblock that has kept autonomous agents out of enterprise offices: trust and liability.

For knowledge workers, Project Arc could finally automate complex, multistep drudgery like server setup, log parsing, and routine administration across multiple local apps. But the barrier to entry remains steep. You will need to be firmly entrenched in ServiceNow’s enterprise stack to get the audit and governance benefits, which keeps smaller shops on the sidelines for now.

For the broader industry, this move puts serious pressure on startups building autonomous desktop agents. Anyone can build a model that clicks buttons on a screen, but very few can hand Fortune 500 CISOs an end-to-end sandbox runtime alongside strict compliance guarantees. By pairing ServiceNow’s operational reach with NVIDIA’s hardware and OpenShell runtime, both giants are turning enterprise governance into their primary moat.

The real test will be reliability. Running multi-step tasks across real operating systems has historically been brittle, and even inside OpenShell, giving AI direct terminal access carries operational risks. Whether enterprises actually trust an agent to work unmonitored will depend on how cleanly Project Arc performs when edge cases crop up in live production.