Source: developer.nvidia.com/blog/build-specialized-agents-on-a-stack-you-control-with-nvidia-nemoclaw
NVIDIA describes NemoClaw as a collection of blueprints rather than a harness or a runtime. Each blueprint combines a tuned open harness, Nemotron open models, NeMo tools for optimization, including NeMo Relay for observing agent execution, and OpenShell, which governs agent access to files, networks, credentials and tools. Developers already building with OpenClaw, Hermes Agent or LangChain Deep Agents can add this surrounding stack without rewriting their agent.
LangChain tuned its Deep Agents harness for Nemotron 3 Ultra by adjusting prompts, tools and middleware, without retraining the model. The company reports near-frontier performance at roughly 10x lower cost per run compared with leading closed models, and the tuned model profile is available on GitHub. The NemoClaw blueprint for that harness packages these optimizations so teams do not have to repeat the tuning work.
An example in the post is a memory-driven Chief of Staff built as a Hermes profile. OpenShell keeps credentials for managed inference and MCP connections outside the sandbox, and NeMo Relay closes trace spans at session boundaries for audit. NemoClaw agents can run in the cloud, in data centers, on DGX Spark and on RTX PCs, and can take delegated subtasks from Claude, Codex or Cursor agents.
Why it matters if you build agents
Teams can run a supported harness on an open Nemotron stack with policy enforcement already included, instead of building that governance layer themselves. OpenShell policies are meant to keep an agent within bounds even when its memory or retrieved context pushes it toward an unintended action. The cost figures come from LangChain's tuning of Deep Agents for Nemotron 3 Ultra, so results on other harnesses or models may differ.
Get stories like this every weekday
One short email with the day's agent tooling news.