The NemoClaw Deep Agents Blueprint represents a collaborative effort between LangChain and NVIDIA to address the complexities of deploying agent systems in production environments. The blueprint focuses on providing enterprises with complete control over their agent stacks, recognizing that agent memory, workflows, and evaluation data represent valuable intellectual property. This approach enables teams to tailor agents to specific workloads, optimizing for quality, cost, and speed.
The core of the blueprint utilizes NVIDIA Nemotron 3 Ultra, a powerful model, alongside LangChain Deep Agents Code (dcode) for the agent harness. The NVIDIA OpenShell runtime provides a secure and governed environment for agent execution, incorporating policies for tool interaction and data access. This layered architecture allows for fine-grained control over the entire agent system.
Evaluations conducted by LangChain demonstrated significant performance gains with Nemotron 3 Ultra and the tuned LangChain Deep Agents harness. The system achieved an aggregate score of 0.86 at a cost of $4.48, substantially lower than the next closest performing model, which cost $43.48 – representing a 10x reduction in inference cost. This improvement stems from careful tuning of tool usage, context management, and intermediate step evaluations.
This optimized stack facilitates larger evaluation suites, enabling more comprehensive testing before deployment and in production. The cost-efficient open architecture supports the comparison of diverse model, harness, and tool variants, facilitating the development of specialized agents for specific domains. The blueprint’s focus on quality, cost, speed, and governance allows companies to tailor agents to their unique requirements, driving innovation and operational efficiency.
“Super agents have arrived,” according to Jensen Huang, Founder and CEO of NVIDIA. The blueprint’s open nature and support from partners like EY and Baseten aim to empower enterprises to build custom AI agents that leverage proprietary data and workflows, fostering a future of AI that is not constrained by one-size-fits-all solutions.
Source: https://www.langchain.com/blog/langchain-and-nvidia-launch-the-nemoclaw-deep-agents-blueprint
