Nvidia has introduced new software, open models and partner integrations for building autonomous AI agents, with a focus on enterprise use in engineering, healthcare, software development and business operations.
The package includes NemoClaw blueprints, Nemotron models, the OpenShell runtime and access to CUDA-X libraries as agent skills. The components are intended to help companies build long-running agents that can orchestrate tasks, retain context, use tools and operate within security and privacy controls.
Several software groups are already using the technology in design and simulation, where verification and modelling tasks can take days or weeks. Cadence, Dassault Systèmes, Siemens and Synopsys are among those using NemoClaw to build autonomous AI engineers that work alongside staff on simulation and verification workflows.
In chip design, Cadence is using OpenShell to secure its ChipStack AI Super Agent, which it describes as a fully autonomous AI engineer for design and verification. Nvidia is the first customer using ChipStack to autonomously verify its chip designs.
Dassault Systèmes is using NemoClaw and OpenShell in its 3DEXPERIENCE agentic platform for long-running agents across design, simulation and manufacturing operations. Siemens is integrating the same software into Fuse EDA AI Agent, designed to plan and orchestrate multi-tool workflows across semiconductor, 3D integrated circuit and printed circuit board system design.
Synopsys is also working with Nvidia on autonomous AI engineers for chip design, focusing on full workflow autonomy. Beyond industrial design, Foxconn is piloting NemoClaw in its Nurabot and CoDoctor platforms for clinical reasoning, documentation and care coordination, while also using Nvidia software in a factory operations agent called MoMClaw.
Model update
Nvidia also unveiled Nemotron 3 Ultra, a 550-billion-parameter mixture-of-experts model aimed at long-running agents in coding, research and enterprise workflows. It says the model offers up to five times faster inference and up to 30% lower cost than open frontier models in the same class.
The model has been post-trained for a range of agent frameworks, including Hermes Agent, LangChain Deep Agents, OpenClaw, OpenHands and OpenCode. Nvidia also announced additional Nemotron models for safety and speech recognition, extending the family beyond general task execution.
Those models are already being used in cybersecurity and operations. CrowdStrike is using Nemotron models in specialised agents that identify, prioritise and remediate vulnerabilities and policy misconfigurations, while Palantir is integrating them into its AI FDE platform to carry out complex tasks in domain-specific and air-gapped enterprise systems.
Security layer
A central part of the announcement is OpenShell, which Nvidia describes as a secure open-source runtime for AI agents. The runtime is designed to set policy and privacy controls as agents gain access to files, tools and memory across longer sessions.
Nvidia is working with Microsoft on a native Windows environment for personal agents using new Windows security primitives alongside OpenShell. According to Nvidia, the controls cover identity, containment, policy and end-to-end security. OpenShell can also route queries to local models based on user privacy settings and mask personal information before requests are sent to cloud models.
Support is also expanding across Linux and enterprise infrastructure. Canonical plans to integrate OpenShell with Ubuntu through supported snaps and rocks, while Red Hat is integrating it into Red Hat AI and contributing to the upstream open-source project.
SAP and ServiceNow have already integrated OpenShell into their products. SAP is embedding it into the Joule Studio runtime in its business AI platform, while ServiceNow is using it to secure Project Arc, its autonomous desktop agent, with policy-based management controls.
Agent skills
Nvidia says its CUDA-X libraries are now available to agents as domain-specific skills. These include cuDF for data processing and analytics, cuOpt for routing and scheduling, AI-Q for research and knowledge workflows, NeMo for optimisation and governance, PhysicsNeMo for scientific simulation, and CUDA-Q for quantum software tasks.
The libraries allow agents to call specialist software functions directly rather than relying only on language model reasoning. Nvidia also released a broader set of open-source physical AI libraries, models and frameworks for robotics, autonomous vehicles and industrial systems.
Jensen Huang, founder and chief executive officer of Nvidia, said: "The world's software leaders are bringing AI agents into the systems where work gets done - showing how AI coworkers help employees think faster and execute complex tasks to solve bigger problems. NVIDIA 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."
NemoClaw is available now, while OpenShell is in early preview.