# NVIDIA, LangChain Introduce NemoClaw Blueprint to Support Enterprise AI Agents > Source: > Published: 2026-07-09 19:59:33+00:00 NVIDIA and LangChain have introduced an open framework created to help enterprises build and manage AI agents while reducing the cost of running them in production. The new release, NemoClaw for LangChain Deep Agents, combines NVIDIA’s Nemotron 3 Ultra open-weight model and NVIDIA’s OpenShell Runtime with LangChain’s Deep Agents framework. Nemotron 3 Ultra serves as the open model layer that enterprises can customize for domain-specific workloads. LangChain Deep Agents provides orchestration capabilities for planning, memory management, tool use and long-running task execution. NVIDIA OpenShell supplies the governed runtime environment, allowing enterprises to define policies controlling how AI agents interact with enterprise systems and data. Rather than retraining Nemotron 3 Ultra, LangChain optimized the surrounding agent framework. The company tuned system prompts, tool descriptions, middleware and execution traces after analyzing performance data from its Deep Agents benchmark. The result is a harness profile that developers can adopt without modifying the underlying model weights. According to benchmark results released by the companies, the tuned configuration achieved an aggregate score of 0.86 in LangChain’s Deep Agents evaluation suite while incurring an inference cost of $4.48 per run. The next highest-performing model in the benchmark reportedly cost $43.48, suggesting approximately a tenfold reduction in inference expense while maintaining comparable task performance. Lower inference costs are important because organizations rely on evaluation throughout the agent lifecycle, testing changes to prompts, workflows, and datasets during deployment. Reducing inference expense makes it more practical to run larger evaluation suites and deploy agents customized for specific business functions without greatly boosting operating costs. The companies claim that an open architecture gives enterprises greater ownership of the intellectual property created around AI deployments. Agent memory, workflow definitions, evaluation datasets, and tuning data all include proprietary business data. By keeping these components under enterprise control, companies can refine agent behavior while avoiding dependence on closed ecosystems. **Beyond the GPU** The announcement expands NVIDIA’s strategy of building an enterprise AI software portfolio beyond its position as the dominant supplier of GPUs. In recent years the company has steadily added AI models, orchestration software, inference platforms and deployment tools built to complement its hardware business. NemoClaw continues that effort by positioning NVIDIA as a provider of the complete software stack needed for enterprise AI agents rather than simply a chip vendor. The infrastructure surrounding the new blueprint is already expanding. Companies including Abridge, Amdocs and Box are integrating specialized AI agents into their platforms, while EY is providing implementation services for organizations deploying NemoClaw-based systems. Hosted access to Nemotron 3 Ultra is available through cloud inference providers including Baseten, Crusoe Cloud, DeepInfra, Fireworks, Nebius and Together AI, allowing enterprises to run the tuned framework without managing underlying infrastructure. For LangChain, the release supports the view that enterprise AI performance depends on optimizing the entire agent system rather than selecting the most advanced foundation model. As organizations deploy autonomous AI systems, tools for governance, and runtime management are becoming ever more important components.