Nvidia published NemoClaw, a set of open blueprints that assemble a production agent stack around a harness the developer already uses. Each blueprint combines four parts: a harness tuned for Nemotron models, the open Nemotron models themselves, NeMo optimization tools including NeMo Relay for tracing agent execution and routing, and the OpenShell runtime that governs an agent's access to files, networks, credentials and tools. Three harnesses are supported at launch - OpenClaw, Hermes Agent from Nous Research, and LangChain Deep Agents - and Nvidia says a blueprint assembles the surrounding stack without requiring the agent to be rewritten. In the LangChain case, the company says the harness was tuned for Nemotron 3 Ultra by adjusting prompts, tools and middleware without retraining the model, reaching near-frontier results at roughly a tenth of the cost per run of leading closed models on the tasks it measured, and that the tuned profile is published on GitHub. A sample recipe keeps managed-inference and MCP credentials outside the agent's sandbox and closes trace spans at session boundaries for audit.
Our takeThe claim worth testing is that a scaffold change - prompts, tool schemas, middleware - moves a model to near-frontier without touching weights. That is really a statement about how much agent performance lives outside the model, and it is checkable: run the same tasks on the tuned and untuned harness against a fixed model and count retries and tokens, not just pass rate. Cost-per-run numbers also need the abandoned runs included, since a harness that stops early looks efficient until the tasks it gave up on are counted.