NVIDIA (NASDAQ: NVDA) announced new software and partnerships to enable enterprises to build autonomous AI agents at its GTC Taipei conference. The company released the NVIDIA Agent Toolkit, which includes NemoClaw blueprints, Nemotron models, and OpenShell secure runtime for developing AI agents that function as digital coworkers.

Major software companies including Cadence, Dassault Systèmes, Siemens and Synopsys are using NVIDIA NemoClaw to create autonomous AI engineers for simulation and verification workflows. These AI agents are designed to compress engineering work that typically takes weeks into hours.

NVIDIA unveiled Nemotron 3 Ultra, a 550-billion-parameter mixture-of-experts model that delivers 5x faster inference and up to 30% lower cost compared to similar open models. The model is specifically designed for long-running agent applications across coding, research and enterprise workflows.

CrowdStrike and Palantir are integrating NVIDIA Nemotron models into their platforms for cybersecurity and operational decision-making applications. The models enable continuous vulnerability identification and complex task execution within enterprise systems.

NVIDIA partnered with Microsoft to develop Windows security primitives and integrate OpenShell runtime for secure agent operation on Windows PCs. Canonical and Red Hat are also integrating OpenShell into their platforms to provide policy and privacy controls for enterprise agent deployment.

The company made CUDA-X libraries accessible to AI agents as domain-specific skills, including cuDF for data processing, cuOpt for optimization problems, and CUDA-Q for quantum computing applications.

NVIDIA NemoClaw is available now, while OpenShell is in early preview. Nemotron 3 Ultra is expected to be available June 4 through various platforms including Hugging Face and build.nvidia.com.