Building an autonomous agent requires four pieces: a reasoning model, an agent harness, domain‑specific tools and skills, and a runtime that governs production access and permissions. NVIDIA’s NemoClaw supplies open blueprints that combine these components so developers can assemble a stack they own, customize, and run.
What an agent harness provides
An agent harness manages how the model uses tools, maintains context, and executes tasks. In practice an agent needs:
- a model that reasons,
- a harness defining the agent’s operation,
- tools and skills to extend capability,
- a runtime to control what the agent can access in production.
Which harnesses NemoClaw supports
NemoClaw blueprints currently support three harnesses:
- OpenClaw: accesses files and applications, writes code, spawns subagents and preserves context across sessions. It runs on your machine or server and supports long‑running sessions, memory, and scheduled tasks.
- Hermes Agent (Nous Research): designed to improve with use; workflows corrected by teams become durable, reusable skills.
- LangChain Deep Agents: a customizable harness for complex, multistep tasks, combining planning, file‑based context management, memory, and delegation to specialized subagents.
Each harness provides the agent’s tools, skills, orchestration and memory.
What NemoClaw adds for full‑stack production agents
NemoClaw is not itself a harness or a runtime. It is a set of open blueprints that assembles a complete open agent stack on top of a harness chosen by the developer. The stack combines tuned models and harnesses with optimization and runtime governance.
Every NemoClaw blueprint brings together four components:
- a tuned open harness (Hermes Agent, LangChain Deep Agents or OpenClaw) optimized to work with NVIDIA Nemotron;
- Nemotron open models;
- NVIDIA NeMo optimization tools, including NeMo Relay for observing and improving agent execution and routing capabilities to improve throughput, cost and reliability;
- NVIDIA OpenShell, the secure runtime that governs access to files, network, credentials and tools.
If you already build agents with OpenClaw, Hermes Agent or LangChain Deep Agents, a NemoClaw blueprint can assemble the surrounding stack without requiring you to rewrite the agent.
Why co‑optimizing model and harness matters
An agent is the combination of a model (which reasons) and a harness (which provides the environment). When the two are matched, the model can focus on tasks. Mismatches can produce repeated context requests, premature stops, or unproductive loops.
As an example, LangChain tuned its Deep Agents harness for Nemotron 3 Ultra by adjusting prompts, tools and middleware rather than retraining the model. From an already strong baseline, that harness tuning delivered near‑frontier performance for Nemotron 3 Ultra at roughly 10× lower cost per run compared with leading closed models.
NemoClaw’s LangChain Deep Agents blueprint packages these optimizations so every developer using LangChain with Nemotron 3 Ultra can benefit without redoing the tuning. The tuned Nemotron 3 Ultra model profile is available through LangChain on GitHub.
When to use a harness vs. a NemoClaw blueprint
- Use a harness (OpenClaw, Hermes Agent or LangChain Deep Agents) to build and experiment with an agent’s skills, tools, memory and orchestration logic.
- Use a NemoClaw blueprint to assemble and operate a specialized agent on an open stack that includes a tuned harness, models, optimization tools, telemetry and OpenShell runtime controls.
Example: a memory‑driven Chief of Staff built on Hermes
The NemoClaw community repository includes a memory‑driven Chief of Staff recipe. The agent keeps a persistent, human‑readable model of people, projects, priorities and working patterns so it does not need to reconstruct context for each task. The recipe is delivered as a deployable Hermes profile and described in a technical design walkthrough.
The example shows the stack working together: OpenShell governs the agent’s filesystem, process and network access and keeps credentials for managed inference and Model Context Protocol (MCP) connections outside the sandbox. This helps ensure the agent remains policy‑bound even if its memory or retrieved context would lead it toward an unintended action. NeMo Relay is natively integrated into the Hermes session, closing trace spans at real session boundaries (new session, reset, expiry) and exporting resulting trajectories for audit and further optimization.
Agent governance
Every NemoClaw reference includes OpenShell, the runtime that adds security and policy enforcement around the harness and tools, governing what actions the model can take. The OpenShell tutorial shows how to create a first policy to prevent data exfiltration to GitHub; NemoClaw references include more sophisticated policies that manage agent connections across multiple internal tools, sites and communication channels.
Deployment options
NemoClaw agents can run in the cloud, in data centers, on NVIDIA DGX Spark, and on NVIDIA RTX PCs and workstations. Different domain agents (for example a clinical agent for a hospital and a simulation agent for electronic design automation) can run on the same stack while using entirely different tools and skills.
NemoClaw agents can also operate alongside frontier agents: Claude, Codex or Cursor agents can delegate domain‑specific subtasks to NemoClaw‑powered specialists running on open models and on owned infrastructure.
Getting started
- Choose a NemoClaw blueprint for your harness and customize tools, models and runtime policies for your workflow.
- Take the "Deploy Secure Agents with NemoClaw DLI" course for an introductory overview.
- Follow the NemoClaw quickstart to deploy an agent from a blueprint and explore the NemoClaw GitHub repository for full examples and additional blueprints.
NemoClaw aims to make it easier to build and run secure production agents by combining tuned harnesses and open models with optimization, telemetry and runtime governance, while keeping control of infrastructure and customization in the hands of developers.
