Not content with just providing the infrastructure for the next generation of artificial intelligence agents, Nvidia Corp. is also providing the tools for developers to build them.
At Nvidia GTC Taipei 2026, concurrent with the Computex conference, the company unveiled the latest iteration of its Agent Toolkit. It’s a comprehensive suite of software, open-source models and blueprints for building powerful, long-running digital coworkers capable of executing complex workflows across business operations, engineering and cybersecurity tasks.
The Nvidia Agent Toolkit is meant to be an open and accessible foundational stack that provides everything developers need to transform powerful frontier models into fully functional AI agents. The suite includes a selection of highly optimized models and a secure runtime environment that attempts to reduce the friction that has caused many companies to hold off from deploying AI agents at scale.
Though large language models have proven themselves to be capable coding assistants and graphics designers, they tend to struggle when attempting to take on more complex business and operational tasks. One of the main challenges for developers is to create a kind of orchestration layer, often called a “harness,” that manages the model’s memory to preserve context across multi-day sessions, enables agents to use third-party tools and collaborate with other agents.
There’s also the security headaches that AI agents create. When organizations give autonomous agents the freedom to access sensitive files, make changes to their application’s code and create their own sub-agents for offloading tasks, this introduces massive security risks that cannot be contained with traditional software enterprise policies.
These are the challenges Nvidia is looking to address, providing developers with a range of open-source building blocks that enable them to create the agentic harness they need. With the Agent Toolkit, developers will be able to safely orchestrate and secure digital coworkers at large scale.
Nvidia Chief Executive Jensen Huang said if the AI revolution everyone imagines is to happen, AI agents must have a way to operate within the systems where business work gets done.
The core of the Agent Toolkit is Nvidia NemoClaw, a new framework that serves as the main blueprint for building agentic orchestration layers. With NemoClaw, developers have access to ready-made templates that structure the way their agents plan, reason, execute and delegate the tasks they’re being asked to do.
“NemoClaw provides enterprise software developers with the open building blocks to create more secure, long-running AI coworkers that amplify human expertise as they reshape how work gets done,” Huang explained.
Another key component is Nvidia’s Nemotron 3 Ultra, a massive new 550 billion-parameter mixture-of-experts model that’s built specifically for long-running autonomous agents. According to Nvidia, it enables frontier-level reasoning across both coding and research workflows, with up to five times faster inference speeds and 30% lower running costs than comparable frontier models in its class.
Nvidia wants developers to run these agents in the OpenShell Secure Runtime, which is a secure container environment that supports custom security and privacy controls. The OpenShell runtime was developed in collaboration with Microsoft Corp., Canonical Ltd. and IBM Corp.’s Red Hat and integrates with native Windows security primitives to ensure AI agents will remain under the full control of their users and avoid dangerous behavior. Using the runtime, developers can intelligently mask sensitive data before sending queries to cloud-based models, and ensure the most sensitive workloads are routed to local hardware only.
Finally, Nvidia introduced a number of CUDA-X libraries as reusable “Agent Skills,” giving Agents access to a range of specialized capabilities without needing to undergo extensive training first. The plug-and-play skills include cuDF, which enables agents to process massive structured datasets rapidly and reason about their findings, and cuOpt, which gives agents the ability to solve complex problems relating to routing, scheduling, supply chain, resource allocation and decision-optimization in real time.
Other skills include AI-Q, which integrates intelligent routing with persistent context and built-in evaluation for enterprise research workflows; NeMo, for accelerating agent optimization, evaluation and governance; PhysicsNeMo, for agents to undertake complex scientific and engineering simulations; and CUDA-Q, which can gives agents the knowledge they need to generate, test and install quantum programs, simulate quantum computing systems and orchestrate quantum applications.
Nvidia shared what a number of early adopters have done with the revamped Agent Toolkit. For instance, the semiconductor design software firm Cadence Design Systems Inc. used OpenShell to deploy a ChipStack AI Super Agent that can automatically verify new chip designs, with Nvidia as the first customer to use that agent. Siemens AG used the toolkit to develop a Fuse EDA agent that can orchestrate multi-tool workflows in printed circuit board design.
In addition to engineering, others are using AI agents to automate cybersecurity workloads. For instance, CrowdStrike Holdings Inc. has developed a number of agents based on Nvidia’s NemoTron 3 Ultra model to continuously identify and remediate security vulnerabilities. Another customer is Palantir Technologies Inc., which has integrated multiple AI models into its Forward Deployed Engineer platform to create autonomous, air-gapped systems that continuously learn from their previous interactions.
Nvidia said the NemoClaw framework is being made available for developers today, while the OpenShell runtime is currently accessible as an early preview. Nemotron 3 Ultra is set to launch on June 4, and will be available as an Nvidia NIM microservice through Hugging Face, ModelScope and OpenRouter, as well as Nvidia’s own Build platform. The CUDA-X agent skills are also available now via the Claude Code marketplace and the Hermes Skills Hub.
Support our mission to keep content open and free by engaging with theCUBE community. Join theCUBE’s Alumni Trust Network, where technology leaders connect, share intelligence and create opportunities.
About SiliconANGLE Media
SiliconANGLE Media is a recognized leader in digital media innovation, uniting breakthrough technology, strategic insights and real-time audience engagement. As the parent company of SiliconANGLE, theCUBE Network, theCUBE Research, CUBE365, theCUBE AI and theCUBE SuperStudios — with flagship locations in Silicon Valley and the New York Stock Exchange — SiliconANGLE Media operates at the intersection of media, technology and AI.
Founded by tech visionaries John Furrier and Dave Vellante, SiliconANGLE Media has built a dynamic ecosystem of industry-leading digital media brands that reach 15+ million elite tech professionals. Our new proprietary theCUBE AI Video Cloud is breaking ground in audience interaction, leveraging theCUBEai.com neural network to help technology companies make data-driven decisions and stay at the forefront of industry conversations.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。