NVIDIA presents the Nemotron model family and the new Nemotron 3 Ultra, a 550B mixture-of-experts model with frontier reasoning, 1M context length, NVFP4 precision for Blackwell GPUs, and 5x faster inference than its predecessor. The talk covers how specialized AI agents can be orchestrated using the Hermes agent harness running on Microsoft Azure Foundry Hosted Agents. A live demo shows Hermes agent reading an email via Outlook, writing code, opening a GitHub PR, and learning reusable skills across sessions. Microsoft Entra ID integration lets enterprises manage agent identities and permissions the same way they manage human employees, while a unified Foundry toolbox consolidates MCP server access, observability, and session isolation for enterprise governance.