Nvidia has launched the latest version of its Agent Toolkit at GTC Taipei 2026, providing developers with open-source tools to build secure, autonomous AI agents capable of executing complex workflows across business, engineering, and cybersecurity.
The suite includes NemoClaw, a new framework for building agentic orchestration layers that manage memory and task delegation over multi-day sessions. It also features Nemotron 3 Ultra, a 550 billion-parameter model offering five times faster inference and 30% lower costs than comparable models.
To address security risks, Nvidia introduced OpenShell Secure Runtime, a container environment with custom privacy controls developed with Microsoft, Canonical, and Red Hat. It masks sensitive data and restricts workloads to local hardware when necessary.
Additionally, Nvidia released CUDA-X libraries as plug-and-play “Agent Skills,” enabling agents to process datasets, optimize supply chains, and run scientific simulations without extensive training.
Early adopters include Cadence Design Systems, which deployed a ChipStack AI Super Agent for chip verification, and CrowdStrike, which uses Nemotron 3 Ultra for cybersecurity vulnerability remediation.