The debate about whether enterprises are ready to deploy AI agents at scale is over. NVIDIA settled it at GTC 2026 in March, when CEO Jensen Huang unveiled the NVIDIA Agent Toolkit alongside commitments from 17 major enterprise software companies — including Adobe, Salesforce, SAP, ServiceNow, and Siemens.

This is not another demo. These are production systems, and the list of partners means the toolkit will reach into virtually every major corporate environment on the planet.

What NVIDIA Actually Released

The Agent Toolkit is an open-source software stack built around four components:

Nemotron — a family of open models optimized for agentic reasoning tasks. These are designed to handle the research-heavy, multi-step work that enterprise agents actually do: reading documents, synthesizing information, preparing outputs for decision-making.

AI-Q — an agentic search blueprint built with LangChain. It uses a hybrid architecture where frontier models handle orchestration and Nemotron models do the detailed research. According to NVIDIA, this approach cuts query costs by more than 50% compared to frontier-only setups, while still topping the DeepResearch Bench leaderboards.

OpenShell — the centerpiece of the toolkit and arguably the most important component for enterprise buyers. It is an open-source runtime that enforces policy-based security, network controls, and privacy guardrails for autonomous agents. In NVIDIA’s terminology, individual agents are called “claws,” and OpenShell is what keeps them operating within defined boundaries.

cuOpt — a skill library for optimization tasks, useful for agents that need to make decisions involving resource allocation, routing, or scheduling.

Why OpenShell Changes the Conversation

Most boardroom conversations about AI agents eventually hit the same wall: how do we let agents operate inside our systems without losing control of what they access or do?

OpenShell is NVIDIA’s answer. It creates isolated sandboxes for each agent, enforces strict policies around data access and network reach, and provides audit trails that compliance and security teams need. NVIDIA is working with Cisco, CrowdStrike, Google, Microsoft Security, and TrendAI to build OpenShell compatibility into their security tools.

That last point matters. If the major enterprise security vendors build around OpenShell, it becomes a de facto standard for how enterprises govern their AI agent deployments. That is a significant shift in how the industry approaches agentic AI risk.

The Partner Ecosystem

The seventeen launch partners cover the major enterprise software categories:

• Business operations: Salesforce, SAP, ServiceNow, Atlassian
• Design and media: Adobe, Dassault Systèmes
• Engineering and manufacturing: Siemens, Cadence, Synopsys
• Security: CrowdStrike, Cisco
• Life sciences: IQVIA
• Infrastructure: Red Hat, Cohesity, Box, Amdocs, Palantir

IQVIA has already deployed more than 150 agents in internal teams and client environments, including at 19 of the top 20 pharmaceutical companies. That is a real deployment number, not a pilot figure.

Salesforce is building a reference architecture where Agentforce agents are orchestrated through Slack, pulling from both on-premises and cloud data environments. Atlassian is integrating Agent Toolkit into Rovo AI across Jira and Confluence. Siemens launched the Fuse EDA AI Agent, which uses Nemotron to autonomously orchestrate workflows in its electronic design automation portfolio.

What This Means for Business

For enterprises already using these platforms: If your organization runs on Salesforce, SAP, or ServiceNow, your vendor is now embedding AI agents directly into your workflows. The toolkit they are using is open and documented. Now is the time to understand what those agents will have access to and what governance you want in place before they go live.

For mid-market businesses: The enterprise adoption signal here is strong. When 17 major vendors converge on a single open standard for agent deployment, it typically means the tooling and support ecosystem around it will mature quickly. What starts as enterprise infrastructure tends to become mid-market infrastructure within 18 to 24 months.

For businesses evaluating AI automation: The AI-Q component is particularly interesting for cost-sensitive deployments. A 50% reduction in query costs for agentic search tasks is meaningful if you are running agents at any volume. The hybrid frontier-plus-open-model architecture is a practical template for building agents that are both capable and affordable to run.

On security and governance: OpenShell addresses the single most common reason enterprise AI agent projects stall: “we do not know what the agent will do.” Policy-based sandboxing with audit trails does not eliminate that concern entirely, but it creates the control surface that security and compliance teams need to approve deployments.

The Bigger Picture

Jensen Huang called this the beginning of “the next industrial revolution in knowledge work.” That framing is consistent with everything NVIDIA has said about the agentic AI opportunity it sees — a $1 trillion market in Huang’s projection, driven by autonomous systems doing cognitive work that previously required human time.

The NVIDIA Agent Toolkit announcement is less about any single product and more about what it signals: the enterprise software ecosystem is converging on open standards for agent deployment. The pilot era is ending. The production era is beginning.

For any business still in the “we are evaluating AI agents” phase, the clock on comfortable evaluation is getting shorter.

---

If this is the kind of problem agents can help with, the free Working With Claude field guide is the practical next step. Thirty-two pages, no fluff. Get the free guide.