NVIDIA Nemotron 3 Ultra achieves leading performance with LangChain Deep Agents harness at lower cost

NVIDIA Nemotron 3 Ultra, when paired with LangChain’s Deep Agents harness, delivers leading performance at a lower cost than top closed models, according to NVIDIA’s Deep Learning Blog. LangChain’s tuning of the Deep Agents harness for Nemotron 3 Ultra achieved the highest accuracy among open models while completing more tasks at higher throughput. The setup also reduced inference cost per run by 10x compared to leading closed models.

Measured against LangChain’s Deep Agents benchmark, Nemotron 3 Ultra achieved business task parity with the highest-scoring closed models without requiring model retraining. All performance gains came from engineering the environment around the model, including system prompts, tool descriptions, and middleware. The tuned profile is available directly through LangChain for developers to use today.

LangChain’s Deep Agents platform, which has over 200 million monthly downloads, enables enterprises to build specialized agents using a fully open stack they can customize, own, and run anywhere. The integration with Nemotron 3 Ultra allows teams to run evaluations continuously, experiment faster, and deploy agents across more business workflows at a fraction of the cost of closed alternatives.

NVIDIA’s NemoClaw for LangChain Deep Agents is an open reference blueprint that packages this work for enterprises. It combines the tuned LangChain Deep Agents harness with the NVIDIA OpenShell secure runtime, providing an end-to-end open stack for building and deploying specialized AI agents. This stack can be customized and run on an enterprise’s own infrastructure or cloud, with governance controlled internally.

Enterprises such as Abridge, Amdocs, and Box are embedding specialized agents into their platforms using this stack, while global systems integrator EY is expanding its NVIDIA implementation capabilities around NemoClaw blueprints. These efforts support clients in customizing, evaluating, and governing specialized agents across high-value workflows.