NVIDIA Nemotron Achieves Benchmark-Leading Performance With LangChain Deep Agents Harness

NVIDIA says LangChain’s Deep Agents harness, tuned for its Nemotron 3 Ultra model, achieved the highest accuracy among open models on LangChain’s Deep Agents benchmark and matched the business-task performance of the top-scoring closed models. According to NVIDIA, the system completed more tasks at higher throughput and cost one-tenth as much per inference run as leading closed models; the gains came from changes to prompts, tool descriptions and middleware rather than model retraining. The tuned profile is available through LangChain, while NVIDIA’s NemoClaw blueprint combines it with the OpenShell secure runtime for enterprises building and governing specialized agents. The results, which come from NVIDIA’s account of LangChain’s benchmark, point to agent-harness engineering and an open, customizable software stack as potential ways to improve performance and reduce operating costs.