ServiceNow opened Knowledge 2026 this week with the announcement everyone in enterprise IT was waiting for and almost nobody covered correctly. The headline product is Project Arc, a joint ServiceNow and Nvidia desktop agent that takes natural-language instructions and executes multi-step work directly on a user’s machine: opening apps, calling APIs, modifying files, navigating between systems. The technically interesting layer underneath is the part to read closely.

Project Arc runs inside a sandboxed runtime called Nvidia OpenShell, with visibility into every file accessed, every command executed, every API called, and every policy enforced. The agent loop is built on top of ServiceNow Action Fabric, the company’s orchestration layer for tool execution. Governance flows through ServiceNow’s AI Control Tower, which is the policy and audit plane it spent the last eighteen months selling enterprise CIOs on. The pitch in the keynote was not “look at this agent doing magic.” It was “here are the logs from the agent doing the magic, here is the policy that blocked the action it should not have done, here is the export you hand to the auditor.”

This is a meaningful shift in posture. A year ago the enterprise agent narrative was a demo of an autonomous AI booking a flight, with applause. The current narrative is closer to “here is the audit trail, here is the rollback, here is the seven-layer permission model you can actually deploy in a regulated environment.” Camunda’s ProcessOS launch last week did the same thing. So did Anthropic’s MCP tunnels and self-hosted sandboxes. The pattern is no longer “AI does cool stuff.” It is “AI does cool stuff, and a paranoid security director is watching.”

The competitive subtext is that ServiceNow has the workflow installed base nobody else can replicate. Salesforce has Agentforce, Microsoft has the Copilot agents, Anthropic has Managed Agents, but ServiceNow already owns the change-management, ticketing, and IT-operations surface area where these agents have to actually do their work. Bundling the agent with the audit plane is the move. The boring infrastructure win in enterprise AI keeps looking less boring as the boring infrastructure keeps eating the demo budget.

servicenownvidiaproject-arcagentsknowledge-2026governanceai-control-toweropenshell