NVIDIA announced its Agent Toolkit at GTC Taipei on June 1, packaging an orchestration framework, a purpose-built open model, and a secure runtime into a single stack for enterprise AI agent development. The announcement positions NVIDIA as more than an inference hardware vendor: it wants to own the software layer where agents plan, execute, and interact with enterprise systems.
Three Components, One Stack
The toolkit has three core pieces.
NemoClaw is an open-source orchestration framework that provides blueprints for building agentic harnesses. It structures how agents plan, reason, execute, and delegate tasks, connecting to popular agent platforms including Hermes Agent, LangChain Deep Agents, OpenClaw, OpenHands, and OpenCode. According to NVIDIA’s announcement, NemoClaw is available to developers starting June 1.
Nemotron 3 Ultra is a 550-billion-parameter mixture-of-experts model post-trained specifically for long-running agentic workloads. NVIDIA claims 5x faster inference and up to 30% lower cost compared with open frontier models in its class. The model is set to launch on June 4 as an NVIDIA NIM microservice, with distribution through Hugging Face, ModelScope, OpenRouter, and NVIDIA’s own Build platform, according to SiliconANGLE.
OpenShell is a secure runtime environment developed in collaboration with Microsoft, Canonical, and Red Hat. It introduces identity management, containment, and policy enforcement for autonomous agents. As SiliconANGLE reported, developers can use OpenShell to mask sensitive data before routing queries to cloud models and enforce that sensitive workloads stay on local hardware. OpenShell is currently available as an early preview.
CUDA-X Libraries as Agent Skills
NVIDIA also opened its CUDA-X libraries as reusable “Agent Skills” that agents can call without additional training. These include cuDF for structured data processing, cuOpt for routing and scheduling optimization, NeMo for model governance, PhysicsNeMo for scientific simulations, and CUDA-Q for quantum computing workflows. The skills are available through the Claude Code marketplace and Hermes Skills Hub, according to SiliconANGLE.
Early Adopters Span Semiconductors, Cybersecurity, and Defense
The adoption list signals where NVIDIA sees the highest-value agent use cases. Cadence is using OpenShell to secure its ChipStack AI Super Agent for autonomous chip design verification, with NVIDIA as the first customer. Siemens integrated NemoClaw and OpenShell into its Fuse EDA AI Agent for multi-tool semiconductor and PCB design workflows. Synopsys is building always-on autonomous AI engineers for chip design. Dassault Systèmes is using NemoClaw to productize agentic capabilities across design, simulation, and manufacturing on its 3DEXPERIENCE platform, according to NVIDIA.
On the security side, CrowdStrike built agents on Nemotron 3 Ultra for continuous vulnerability identification and remediation. Palantir integrated multiple AI models into its Forward Deployed Engineer platform to create autonomous, air-gapped systems that learn from previous interactions, as SiliconANGLE reported.
Foxconn is piloting NemoClaw for healthcare (clinical reasoning via its Nurabot and CoDoctor platforms) and factory operations (real-time sensor and machine data analysis), according to NVIDIA.
The Infrastructure Play
“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,” Jensen Huang said in the announcement.
The strategic logic is straightforward: NVIDIA already dominates inference compute. By providing the orchestration framework (NemoClaw), the model (Nemotron), and the runtime (OpenShell), it is building a vertically integrated agent stack where every layer runs best on NVIDIA hardware. Enterprise teams that adopt these tools become more dependent on NVIDIA silicon, not less. Whether that consolidation benefits or constrains the broader agent ecosystem will depend on how open-source competitors respond.