ServiceNow and NVIDIA introduced Project Arc at ServiceNow Knowledge 2026 (May 5-7) — a long-running, self-evolving autonomous desktop agent that knowledge workers run locally, with full file-system, terminal, and application access, governed by ServiceNow’s Action Fabric and sandboxed by NVIDIA’s open-source OpenShell runtime. ServiceNow Project Arc is the most technically ambitious enterprise desktop agent to ship to date — designed for developers, IT teams, and administrators handling complex multi-step work that traditional automation cannot. The combination of local desktop access plus enterprise governance is what makes Project Arc consequential — desktop agents previously had to choose between capability (full local access, weak governance) or governance (restricted scope, weaker capability). Project Arc delivers both through an architecture worth understanding.
What’s actually new
Project Arc has three architectural pillars that distinguish it from prior desktop-agent approaches. First, NVIDIA OpenShell — an open-source secure runtime for developing and deploying autonomous agents in sandboxed, policy-governed environments. OpenShell handles the security boundary between the agent’s local-machine access and the enterprise’s policy enforcement, which has been the chronic gap in earlier desktop-agent products. Second, ServiceNow Action Fabric integration — every action the agent takes flows through the ServiceNow AI Platform’s governance, auditability, and workflow intelligence layer. The agent doesn’t operate in isolation from enterprise systems; it operates as a governed extension of them. Third, self-evolution — the agent learns from its work over time without losing the governance constraints. The pattern is similar to Anthropic’s Dreaming feature for managed agents but applied to desktop autonomy.
The use cases Project Arc targets are concrete. Developer workflows: setting up environments, running tests, debugging issues, opening pull requests. IT operations: incident triage, system health checks, configuration changes within policy. Administrator tasks: user provisioning, policy updates, audit-log review. Each of these involves multi-step work spanning multiple applications that traditional automation tools cannot reliably handle but that human knowledge workers do routinely. Project Arc is positioned to handle these workflows with appropriate human supervision.
The benchmarking comes alongside the product release. NOWAI-Bench — including EnterpriseOps-Gym and EVA-Bench — is available as an open-source release. The benchmarks measure agent performance on enterprise operational tasks rather than the academic benchmarks (SWE-bench, GAIA, etc.) that have dominated agent evaluation. The pattern of releasing domain-specific benchmarks alongside products is meaningful — it lets the community evaluate Project Arc and competing approaches on workloads that match real enterprise use.
The integration with ServiceNow’s broader AI Control Tower extends governance from desktops to data centers. The AI Control Tower integration with NVIDIA Enterprise AI Factory validated design is generally available, meaning organizations can manage agentic workloads from local-employee desktops through to production data center workloads under unified governance. The architectural ambition is substantial — it’s the agentic-AI equivalent of the unified observability that monitoring vendors built across infrastructure layers.
Pricing and availability are tiered. Project Arc is initially available to ServiceNow customers in early access, with broader availability rolling out through Q3-Q4 2026. NOWAI-Bench is open-source available now. The NVIDIA OpenShell runtime is available on GitHub under permissive open-source license — meaningful because it lets organizations beyond ServiceNow’s customer base build their own desktop agents on the OpenShell foundation.
Why it matters
- Desktop autonomous agents finally have credible enterprise governance. Prior desktop automation tools (RPA tools, browser-only agents, OS-level agents) had to choose between capability and control. Project Arc’s architecture combining OpenShell sandbox plus Action Fabric governance addresses the fundamental constraint that has limited enterprise deployment.
- NVIDIA OpenShell as open source matters strategically. A vendor-neutral runtime for sandboxed agent execution lets the broader ecosystem build secure desktop agents without rebuilding the foundation. Expect rapid adoption across other agent frameworks through 2026-2027.
- The NVIDIA-ServiceNow partnership extends both companies’ enterprise positioning. NVIDIA gets reach into enterprise governance scenarios beyond pure infrastructure. ServiceNow gets compute and runtime capability beyond what software-only platforms could offer. The combination is more capable than either alone.
- NOWAI-Bench fills a real gap in agent evaluation. Enterprise operational tasks differ from academic benchmark tasks. Open-source benchmarks specifically targeted at enterprise scenarios let the agent ecosystem improve on dimensions that matter for production deployment rather than benchmark performance that doesn’t translate.
- The unified governance from desktop to data center is the long-term play. AI Control Tower spanning local employee work and production AI factories produces consistent governance regardless of where AI workloads run. The architectural pattern will likely become the enterprise default through 2027-2028.
- The competitive pressure on Microsoft, Google, and other agent platform vendors increases. Microsoft Copilot Studio, Google’s agent platform, and OpenAI‘s Agent SDK all face a more capable enterprise-governance challenger. Expect competitive responses through Q3-Q4 2026.
How to use ServiceNow Project Arc today
Three steps put a ServiceNow customer on Project Arc.
- Request early access through your ServiceNow account team. Project Arc is in early access through Q3 2026. Existing ServiceNow customers can request enrollment; the rollout is staged across customer segments. Enterprise and strategic accounts get priority access.
- Plan the deployment scope. Project Arc’s value comes from handling multi-step workflows that span applications. Identify the highest-value workflows in your developer, IT operations, or administrator teams. Start with one or two workflows where the value is clear and the risk is bounded.
- Configure governance through ServiceNow Action Fabric. The Action Fabric layer defines what actions the agent can take, what data it can access, what approvals are required for consequential actions, and what audit logging is captured. Configure conservatively at first; expand scope as confidence builds.
For developers building on the NVIDIA OpenShell runtime directly, the open-source library handles the sandboxed agent execution layer:
MASK12
The OpenShell library handles the sandboxing primitives, policy enforcement, and audit logging. The agent model itself is pluggable — Claude, GPT, or other foundation models can drive the agent’s reasoning while OpenShell maintains the security boundary.
How it compares
The desktop and enterprise agent landscape in mid-2026 has multiple approaches with different tradeoffs. The table below summarizes the leaders along the dimensions that matter for enterprise procurement.
| Product | Surface | Governance | Self-evolution | Best fit |
|---|---|---|---|---|
| ServiceNow Project Arc + NVIDIA OpenShell | Desktop with full local access | Action Fabric + AI Control Tower | Self-evolving | ServiceNow-shop enterprises |
| Microsoft Copilot Studio + Wave 3 | M365 + connected apps | Microsoft Entra + Purview | Limited self-evolution | Microsoft-shop enterprises |
| Anthropic Claude Managed Agents (with Dreaming) | Web + connected services | Workspace-style admin | Dreaming self-improvement | Customer experience and high-stakes workflows |
| OpenAI Agent SDK + Operator | Browser + tool calls | Custom integration | Limited | OpenAI-stack deployments |
| Google Gemini Agent | Workspace + Search | Workspace admin | Limited | Google Cloud / Workspace shops |
| UiPath / Automation Anywhere (RPA + AI) | Desktop + applications | RPA platform | Limited; rule-driven | RPA-mature operations |
Two takeaways. First, Project Arc’s combination of full desktop access plus enterprise governance is genuinely distinctive. Microsoft’s Copilot operates within Microsoft’s surfaces; Anthropic’s managed agents operate in web contexts; OpenAI’s Operator is browser-focused. Project Arc handles the local-machine work that none of these directly address. Second, the governance integration is what differentiates Project Arc from RPA-plus-AI approaches. RPA platforms have desktop access but their AI integration is bolted on rather than architectural. Project Arc was designed agent-first with governance as a foundational concept rather than an afterthought.
What’s next
Three things to watch over the next two quarters. First, Project Arc’s broader availability rollout. Early access is open now; broader availability is targeted for Q3-Q4 2026. The pace of customer adoption and the customer-reported outcomes will determine how quickly Project Arc reaches market dominance in the desktop-agent category. Second, OpenShell ecosystem development. As an open-source runtime, OpenShell could attract integrations beyond ServiceNow — possibly competing agent platforms building on the same secure runtime. Microsoft, Google, and others may either contribute to OpenShell or develop competing runtime standards. Third, the regulatory dimension. Desktop agents with full local-machine access raise specific compliance concerns — privacy, data handling, audit requirements. The regulatory environment will likely sharpen through 2027 as Project Arc and similar products reach broader deployment.
The longer-term implication is that the desktop-agent category is finally maturing. Earlier attempts at desktop AI agents (X.ai’s Lindy, the original Microsoft Recall, various RPA-AI hybrids) faced architectural constraints that limited adoption. Project Arc’s architecture addresses the most fundamental constraints. The 2027 enterprise software landscape will likely include desktop autonomous agents as standard capability rather than experimental technology, which will reshape how enterprise software vendors design their products.
Frequently Asked Questions
What’s the difference between Project Arc and existing RPA tools?
Architecture and capability. RPA tools (UiPath, Automation Anywhere, Blue Prism) were designed for rule-based automation and have AI bolted on. Project Arc was designed agent-first — the AI reasoning is the core capability, with sandboxing and governance as architectural primitives rather than retrofits. The result is meaningfully better handling of multi-step work that doesn’t fit predetermined rules.
Can I use Project Arc without ServiceNow?
Project Arc itself is part of ServiceNow’s platform. The underlying NVIDIA OpenShell runtime is open source and can be used independently to build sandboxed agents without ServiceNow. Organizations not on ServiceNow can build similar capability on OpenShell but must add the governance and integration layers ServiceNow provides natively.
How does the security model work for desktop agents with full local access?
Sandboxing through OpenShell defines what the agent can and cannot access. Policy enforcement at runtime prevents actions outside the allowed scope. Audit logging captures every action for review. Approval gates require human confirmation for consequential actions. The combined controls make full local access tractable for enterprise deployment in ways that earlier desktop agent approaches could not achieve.
What kinds of tasks should I delegate to Project Arc?
Multi-step routine work that doesn’t fit predetermined automation rules. Examples: setting up developer environments, running diagnostic workflows, processing administrative tasks across multiple applications, executing complex configuration changes within policy. Tasks requiring substantial judgment about business context or strategy remain primarily human work; tasks that follow established patterns benefit from Project Arc.
How does Project Arc compare to Microsoft Copilot’s automation capabilities?
Microsoft Copilot operates within Microsoft 365 surfaces with strong integration into Microsoft applications. Project Arc operates at the OS and application level with broader local access. For Microsoft-shop organizations whose work happens primarily in Microsoft 365, Copilot may be sufficient. For organizations whose work spans diverse applications and requires deeper local access, Project Arc addresses scenarios Copilot cannot reach. Many organizations will use both for their respective strengths.
What’s the long-term implication of unified governance from desktop to data center?
It enables consistent AI policy enforcement regardless of where AI workloads run. An employee’s local agent operates under the same governance framework as production AI services in data centers. The unification produces simpler audit trails, more consistent policy application, and better visibility for IT and security teams. The architectural pattern is likely to become the enterprise default for AI governance through 2027-2028.