Autonomous AI agents are moving from experiment to enterprise infrastructure faster than most organizations anticipated. The question is no longer whether companies will deploy AI agents but what software foundation those agents will run on, how secure they will be, and whether businesses can actually control what they do. NVIDIA has an answer it is pushing hard, and at GTC this week the company made clear just how broadly it intends to embed itself in the enterprise agent ecosystem.
The company announced a significant expansion of the NVIDIA Agent Toolkit, an open source software platform for building and running autonomous AI agents. The headline addition is NVIDIA OpenShell, a new open source runtime that enforces security, network, and privacy guardrails for agents operating inside enterprise environments. Alongside it, NVIDIA unveiled the AI-Q Blueprint, an open agent framework for enterprise search and research that is already sitting at the top of accuracy leaderboards for that category. And the list of enterprise software companies integrating Agent Toolkit into their platforms reads like a who’s who of the business software industry.
What the Agent Toolkit Is and Why It Matters
NVIDIA Agent Toolkit is a collection of open source components designed to give enterprises and developers the building blocks for creating AI agents that can autonomously determine how to complete assigned tasks. The toolkit includes open models in the form of NVIDIA Nemotron, open agents through the AI-Q Blueprint, open skills via NVIDIA cuOpt, and now the OpenShell runtime for security and privacy enforcement.
The framing Jensen Huang used at the announcement is worth noting. He pointed to Claude Code and OpenClaw as having sparked what he called the agent inflection point, the moment when AI moved beyond generating content and reasoning about problems into actually taking action. His argument is that enterprises are about to deploy teams of frontier, specialized, and custom-built agents, and that the software industry is on the brink of a major transformation as a result. Whether or not that timeline plays out exactly as Huang describes, the scale of enterprise adoption already forming around Agent Toolkit suggests the direction is real.
OpenShell: Security Guardrails for Autonomous Agents
The most important new addition to the Agent Toolkit from an enterprise adoption standpoint is OpenShell. When an AI agent has the ability to access files, run code, call APIs, and interact with external services autonomously, the security and privacy implications are significant. Without guardrails, agents can inadvertently expose sensitive data, take actions outside their intended scope, or create vulnerabilities that malicious actors can exploit.
OpenShell addresses this by creating a controlled runtime environment in which agents operate. It enforces policy-based security controls, manages network access, and applies privacy guardrails that determine what agents can do and where they can send data. The goal is to give agents the access they need to be genuinely productive while keeping that access within boundaries the enterprise controls.
NVIDIA is not building the security layer alone. The company is collaborating with Cisco, CrowdStrike, Google, Microsoft Security, and TrendAI to ensure OpenShell is compatible with their respective cybersecurity and AI security tools. That kind of integration with established enterprise security platforms is important because it means OpenShell can fit into existing security workflows rather than requiring organizations to build a parallel security infrastructure just for their agents.
The AI-Q Blueprint and What It Does on Leaderboards
The NVIDIA AI-Q Blueprint is an open agent framework built in collaboration with LangChain that enables developers to create custom AI agents capable of perceiving, reasoning, and acting on enterprise knowledge. An agent built on AI-Q can automatically choose the right data sources for a given query, determine how deeply to analyze the information, and deliver context-aware answers along with a built-in explanation of how each answer was produced.
The architecture is designed to be practical from a cost standpoint as well as a capability standpoint. AI-Q uses frontier models for orchestration and NVIDIA Nemotron open models for the research tasks, a hybrid approach that NVIDIA says can cut query costs by more than 50 percent while maintaining high accuracy. The company used the AI-Q Blueprint to develop the top-ranked AI agent on both the DeepResearch Bench and DeepResearch Bench II leaderboards, which are benchmarks specifically designed to evaluate the accuracy of AI research agents.
LangChain, whose open source frameworks have been downloaded over a billion times, is integrating AI-Q, OpenShell, and Nemotron models into its deep agent library for building and running advanced enterprise AI agents at scale. That integration gives the Agent Toolkit immediate reach into one of the most widely used agent development ecosystems in the world.
The Enterprise Software Lineup
The breadth of enterprise software companies working with Agent Toolkit is one of the most significant signals about where this platform is heading. The list covers nearly every major category of business software.
Adobe is adopting Agent Toolkit as the foundation for running hybrid, long-running agents for creativity, productivity, and marketing in a more secure and cost-efficient environment. Atlassian is working with Agent Toolkit and OpenShell as it evolves its Rovo AI strategy for tools like Jira and Confluence. Box is using Agent Toolkit to enable enterprise agents that can securely execute long-running business processes using the Box file system.
Salesforce is integrating Nemotron models to help customers build and deploy Agentforce agents for service, sales, and marketing tasks. The collaboration also introduces a reference architecture in which employees can use Slack as the primary interface for interacting with Agentforce agents that participate directly in business workflows and can pull from both on-premises and cloud data stores. SAP is using NVIDIA NeMo to enable AI agents through Joule Studio on its Business Technology Platform, giving customers and partners the ability to design agents tailored to their specific business needs.
ServiceNow’s Autonomous Workforce of AI Specialists is built on the ServiceNow AI Platform and leverages Agent Toolkit, the AI-Q Blueprint, and a combination of closed and open models including Nemotron and ServiceNow’s own Apriel models. Red Hat is integrating Agent Toolkit into its Red Hat AI Factory with NVIDIA to provide an enterprise-ready platform for building more secure, autonomous agents.
On the security side, CrowdStrike has unveiled a Secure-by-Design AI Blueprint that embeds Falcon platform protection directly into NVIDIA AI agent architectures including those built on AI-Q and OpenShell. The company is also using Nemotron reasoning models and NeMo Data Designer to power AI agents for investigative workflows within its managed detection and response offerings. Cisco AI Defense is providing AI security protection for OpenShell, adding controls and guardrails to govern agent and claw actions.
In specialized industries, Amdocs is using AI-Q and Nemotron to power its Cognitive Core agent platform, which continuously monitors customer interactions and billing data to proactively identify and resolve issues before customers are impacted. IQVIA is integrating Nemotron and other Agent Toolkit components with its unified agentic AI platform, and has already deployed more than 150 agents across internal teams and client environments including 19 of the top 20 pharmaceutical companies. Palantir is using Nemotron to develop AI agents running on its sovereign AI Operating System Reference Architecture.
For the semiconductor and hardware design industries, Cadence is using Agent Toolkit and Nemotron with its ChipStack AI SuperAgent to help engineers design and verify more complex semiconductor designs. Siemens is launching the Fuse EDA AI Agent, which uses Nemotron to autonomously orchestrate workflows across its electronic design automation portfolio for semiconductor and printed circuit board design from conception through manufacturing sign-off. Synopsys is building a multi-agent framework for semiconductor and systems design using Nemotron and NeMo Agent Toolkit.
Dassault Systèmes is exploring Agent Toolkit and Nemotron for its role-based AI agents called Virtual Companions on the 3DEXPERIENCE agentic platform. Cohesity is adding OpenShell support and expanding its Gaia AI platform to support more advanced agentic workflows with AI-Q.
Where Developers Can Access It
NVIDIA Agent Toolkit and OpenShell are available on build.nvidia.com today. Developers can run them on inference providers and NVIDIA Cloud Partners including Baseten, Bitdeer AI, CoreWeave, DeepInfra, DigitalOcean, Fireworks, GMI Cloud, Lightning, Together AI, and Vultr. OpenShell can also be used with LangChain and downloaded from GitHub to run locally on NVIDIA GeForce RTX PCs and laptops, RTX-powered workstations, and NVIDIA DGX Station and DGX Spark supercomputers from Altos Computing, ASUS, Dell Technologies, GIGABYTE, HP, Lenovo, MSI, and Supermicro.
Enterprises building and running agents on AI factory infrastructure can do so through Amazon Web Services, Google Cloud, Microsoft Azure, Microsoft Security, and Oracle Cloud Infrastructure, as well as on servers from Cisco, Dell Technologies, HPE, Lenovo, and Supermicro.
The Bigger Picture
NVIDIA’s Agent Toolkit strategy is, at its core, an attempt to establish the company as the default infrastructure layer for enterprise AI agents the same way it became the default infrastructure layer for AI training and inference. By open sourcing the core components, building security and privacy features that enterprise IT teams require, and integrating with the software platforms that enterprises already run their businesses on, NVIDIA is creating a set of switching costs that compound over time.
The risk is that the enterprise software companies in this ecosystem build enough of their own agent infrastructure that they reduce their dependence on NVIDIA’s software layer over time. But the depth of the current integrations, and the speed with which the ecosystem is forming, suggests NVIDIA has moved early enough and broadly enough to make Agent Toolkit genuinely foundational for a wide range of enterprise agent deployments in the near term.