NVIDIA releases Agent Toolkit for running enterprise AI agents safely

NVIDIA has introduced Agent Toolkit, an open source stack for companies that want to run AI agents in real internal systems, not just in demos. At GTC 2026 in San Jose, the company showed how it intends to address enterprise customers' biggest fear: losing control over data, access and agent actions.

Why the market is stalling

Corporate customers settled the question of whether they need AI agents long ago. The question now is different: what happens when an agent gets access to a CRM, internal documents, a service desk, billing or production processes. When an ordinary chatbot gets an answer wrong, it is unpleasant. When an autonomous agent launches an action on its own, the error becomes an incident, a leak or a direct cost. That is why the mass move from pilots to production is slower than model vendors promised.

NVIDIA is betting not on yet another "smart assistant" but on an infrastructure layer meant to keep agents within set limits. The company says outright that businesses need guardrails: access rules, network restrictions, privacy controls and clear accountability for the system's actions. Until now, all of this was often assembled by hand for each specific case, which scaled poorly. Agent Toolkit is pitched as an attempt to standardize this part of the market and remove the main barrier to adoption.

What the stack includes

The platform's key element is OpenShell, an open source runtime for autonomous agents. In NVIDIA's terminology the agents themselves are called claws, and OpenShell makes sure they work according to company policy rather than "however they end up working." It sets security, network and privacy constraints and is meant to let an agent do useful work without gaining excess permissions. This is an important shift: the emphasis moves from the model itself to the execution environment and control over behavior.

• OpenShell: an execution environment with security policies and network and privacy restrictions
• AI-Q: a blueprint for agentic search and work with corporate knowledge, built together with LangChain
• Nemotron: NVIDIA's open models for the research and analysis stages inside the agent pipeline
• Built-in response evaluation: a system that shows exactly how the agent arrived at its result

Agent Toolkit also includes AI-Q, an agentic search blueprint built with LangChain. NVIDIA uses a hybrid scheme: frontier models handle orchestration, while open Nemotron models take on the research part. According to the company, this approach can cut query costs by more than 50% while still delivering results at the level of the top entries on DeepResearch Bench and DeepResearch Bench II. For enterprises this is not a cosmetic metric: query economics often break scaling after a successful pilot.

"Claude

Code and OpenClaw have triggered a turning point for agentic AI: it is moving beyond generation and reasoning to action," said Jensen Huang.

So that the story does not look purely theoretical, NVIDIA immediately tied OpenShell to the ecosystem of security vendors. The company announced compatibility and cooperation with Cisco, CrowdStrike, Google, Microsoft Security and TrendAI. The logic is simple: if agents are going to work inside corporate perimeters, they must fit into existing protection tools rather than ask businesses to build a separate world just for AI.

Who has signed on

NVIDIA immediately assembled an impressive list of partners around Agent Toolkit: Adobe, Atlassian, SAP, Salesforce, ServiceNow, Siemens, Cisco, CrowdStrike, Red Hat, Box, Cohesity, Cadence, Dassault Systèmes, IQVIA and Synopsys. This is an important signal to the market: this is not an experimental SDK for enthusiast developers but an attempt to become the base layer under enterprise software. NVIDIA wants its stack to live not next to enterprise software but beneath it, as the standard layer for agentic scenarios.

There are also more concrete examples. Salesforce is building a reference architecture with NVIDIA in which employees use Slack as the main interface and orchestration layer for Agentforce agents that work with data both on-prem and in the cloud. Atlassian is integrating Agent Toolkit into its Rovo strategy for Jira and Confluence. ServiceNow is using it to develop an "autonomous workforce" of AI specialists. Siemens launched Fuse EDA AI Agent to automate workflows in electronics, and IQVIA has already deployed more than 150 agents across internal teams and at clients, including 19 of the 20 largest pharmaceutical companies.

The toolkit is already available at build.nvidia.com and is supported on AWS, Google Cloud, Microsoft Azure and Oracle Cloud Infrastructure.

What this means

NVIDIA is trying to claim the most lucrative tier of the new market: not just selling chips or models, but controlling the runtime, security and economics of enterprise AI agents. If the OpenShell and AI-Q approach really does simplify launching agents in production, the company has a chance to become for the agentic enterprise what Kubernetes became for containers: not the only player, but the layer that nearly every serious deployment passes through.