What happened
On October 30, 2026, NVIDIA’s developer blog posted Build Specialized Agents on a Stack You Control with NVIDIA NemoClaw. NemoClaw is described as open blueprints, not a competing harness: each blueprint packages a tuned open harness (OpenClaw, Hermes Agent, or LangChain Deep Agents), Nemotron open models, NeMo optimization tools including NeMo Relay, and NVIDIA OpenShell as the secure runtime that governs files, networks, credentials, and tools. NVIDIA says LangChain tuned its Deep Agents harness for Nemotron 3 Ultra by adjusting prompts, tools, and middleware without retraining the model, and that the tune brought Nemotron 3 Ultra to near-frontier performance at approximately 10x lower cost per run compared with leading closed models. The tuned model profile is said to be available through LangChain on GitHub. Agents can run in cloud, data center, DGX Spark, and RTX PCs, and can sit beside frontier agents such as Claude, Codex, or Cursor.
Why it matters
A team comparing agent releases has to ask which piece moved. NemoClaw’s headline number is a harness-and-model co-tune cost claim, not a new frontier weight file. OpenShell is the governance layer NVIDIA puts around that stack.
NVIDIA links a Deploy Secure Agents with NemoClaw DLI course and a NemoClaw GitHub repository for blueprints.
OpenShell policies in the references include preventing data exfiltration to GitHub as a starter example.
A community Chief of Staff recipe on Hermes shows memory, OpenShell, and NeMo Relay used together.