NVIDIA and LangChain have unveiled an open enterprise AI agent stack that lets organisations self-host and govern production AI agents while reducing inference costs and avoiding proprietary vendor lock-in.
NVIDIA and LangChain have introduced the NemoClaw for LangChain Deep Agents blueprint, an open reference architecture that enables enterprises to build, govern and deploy production AI agents on their own infrastructure. Combining the open-weight Nemotron 3 Ultra model, the MIT-licensed LangChain Deep Agents framework and NVIDIA OpenShell runtime, the stack gives organisations greater transparency, control and portability while reducing dependence on proprietary AI APIs.
According to LangChain’s benchmark, the blueprint achieved a 0.86 score on its Deep Agents evaluation suite at a cost of $4.48 per completed task, compared with $43.48 for the nearest competing model, representing roughly a 10x reduction in inference cost.
Nemotron 3 Ultra is released under the Linux Foundation’s OpenMDW-1.1 licence, with NVIDIA publishing the model weights, training data, training recipes and fine-tuning code. This level of openness exceeds most open-weight releases and supports transparency requirements under the EU AI Act. LangChain Deep Agents provides planning, memory, tool calling and multi-step execution, while OpenShell adds sandboxed execution, governance controls and policy enforcement to improve security.
“The way to build better agents is to keep improving the system around the model. Memory, tool use, evaluation, and model behavior compound when teams can tune them together,” said Harrison Chase, Co-founder and CEO of LangChain.
The companies say the architecture addresses one of enterprise AI’s biggest barriers as agentic workloads require significantly more model calls than conventional chatbots. However, the reported benchmark is based on LangChain’s proprietary 127-example evaluation and has not been independently verified. Enterprises are therefore advised to validate cost and performance against their own workloads before deployment.
