Project Arc can write code and complete multistep tasks while running inside NVIDIA's sandboxed OpenShell runtime
Bloomberg / Getty Images
ServiceNow $NOW and NVIDIA introduced Project Arc, an enterprise autonomous desktop agent designed to complete complex, multistep work across enterprise tools and systems without requiring pre-built workflows, the companies said at ServiceNow's Knowledge 2026 conference in Las Vegas.
NVIDIA OpenShell, an open-source sandboxed runtime environment, serves as the execution layer for Project Arc, using policy-based management to contain autonomous activity and maintain an auditable record of agent behavior. Governance comes from ServiceNow's AI Control Tower, which enforces policies and maintains logs covering which files the agent accessed, which commands it ran, and which APIs it invoked, the company said.
The agent connects to ServiceNow's Action Fabric, which gives it access to company workflows, systems, and operational history. It can be accessed through a desktop app, enterprise collaboration tools, or email, ServiceNow said.
Jon Sigler, executive vice president and general manager of AI Platform at ServiceNow, described the problem the product addresses. "We went and built our own AI agent for the desktop that can write code that can run these long-running processes in the background, but we needed a way to secure that," Sigler told Yahoo Finance. "And in comes Nvidia $NVDA … they have something called OpenShell, and that allowed us to run our agent in their sandbox, preventing it from doing things that it shouldn't be doing."
NVIDIA's Justin Boitano, who leads enterprise AI platforms, described OpenShell to Yahoo Finance as "a safe runtime that can run any agent." He added that security professionals have coined the term "lethal trifecta" for agents deployed without such protections, pointing to three compounding risks: they hold access to sensitive private data, they can encounter untrusted content, and they retain the ability to reach systems beyond the desktop.
The partnership expansion also extends ServiceNow's AI Control Tower into data center infrastructure. The integration with the NVIDIA Enterprise AI Factory validated design, previewed at NVIDIA's GTC conference in March, is now complete, ServiceNow said. It gives organizations managing large-scale AI workloads a governance layer covering model discovery, compliance monitoring, real-time observability, and cost management.
The companies also released NOWAI-Bench, an open benchmarking suite comprising two frameworks: EnterpriseOps-Gym, covering multistep agentic evaluation across IT service management, customer service, and HR workflows, and EVA-Bench, a voice agent evaluation framework. Both are available as open-source releases, and NVIDIA is integrating them into its NeMo Gym library. NVIDIA has already evaluated its Nemotron 3 Super model on EnterpriseOps-Gym, with results the companies say show the model ranks first among open-weight models on the benchmark.
Project Arc is available as an early preview. The AI Control Tower integration with the NVIDIA Enterprise AI Factory is generally available.
The partnership has deepened steadily over the past year. At NVIDIA's GTC 2026 earlier this year, Huang outlined NVIDIA's strategy for pushing AI agents beyond chatbots and into daily enterprise work, including the rollout of OpenShell as a secure runtime for autonomous agents. ServiceNow and NVIDIA had previously collaborated on the Apriel Nemotron open model family and on specialized AI agents for enterprise workflows.
Join 500,000+ readers who start their day with Quartz.
By subscribing, you agree to our Terms of Service and Privacy Policy.
