NVIDIA’s Nemotron 3 Ultra has achieved benchmark-leading performance in AI agent orchestration by utilizing LangChain’s Deep Agents harness, according to recent performance data. By optimizing the environment surrounding the model rather than retraining the underlying architecture, the integration allows for higher throughput and lower inference costs compared to leading closed models. This development marks a shift in enterprise AI, where performance gains are increasingly derived from engineering system-level interactions, such as memory management and tool use, rather than traditional model fine-tuning.
The collaboration between NVIDIA and LangChain centers on the Deep Agents benchmark, a suite designed to evaluate how AI agents execute multi-step tasks within business workflows. According to the companies, this approach has enabled Nemotron 3 Ultra to reach business task parity with the highest-scoring closed models. By focusing on the “harness”—the middleware, prompts, and tool descriptions that govern how a model interacts with its environment—developers can now achieve high-performance results at a fraction of the cost. This shift effectively lowers the barrier for enterprises aiming to deploy specialized agents that can perform autonomous actions within core business systems.
Engineering the Environment for Agent Performance
The methodology behind this performance leap prioritizes “harness engineering” over traditional model retraining. When LangChain evaluated Nemotron 3 Ultra against its public Deep Agents benchmark, the team performed a granular analysis of execution traces to identify specific points of failure. Instead of modifying the base model, developers adjusted the surrounding system, including middleware and system prompts. This strategy allows teams to iterate rapidly, as the model itself remains static while the surrounding architecture is continuously improved to handle increasingly complex tasks.
Harrison Chase, cofounder and CEO of LangChain, noted that the evolution of AI agents depends on the system architecture. “The way to build better agents is to keep improving the system around the model,” Chase said. “Memory, tool use, evaluation and model behavior compound when teams can tune them together. Our work with NVIDIA shows that enterprises can get strong performance from an open stack while keeping control over the agent systems they are building.”
Building a Fully Open Enterprise AI Stack
For organizations seeking to maintain control over their data and infrastructure, the integration introduces the NVIDIA NemoClaw for LangChain Deep Agents. This open reference blueprint serves as a framework for enterprises to build and govern their own specialized AI systems. By combining the LangChain Deep Agents code with the NVIDIA OpenShell secure runtime, companies can execute agent actions within their own cloud environments or on-premises infrastructure. This “open stack” model is designed to address concerns regarding data sovereignty and governance, particularly as AI agents transition from simple Q&A assistants to tools that actively manage business operations.
The practical application of this technology is already underway with several industry partners. Companies including Abridge, Amdocs, and Box are currently embedding specialized agents into their platforms. Additionally, global systems integrator EY is expanding its implementation capabilities to support the NemoClaw blueprints, assisting clients in customizing and governing agents for high-value workflows. This collaborative ecosystem highlights a broader industry trend where enterprises are moving toward customizable, proprietary agent systems rather than relying exclusively on black-box, closed-source services.
Deployment and Accessibility for Developers
Developers can access the tuned Nemotron 3 Ultra model profile directly through the LangChain platform, enabling immediate integration into existing projects. To facilitate production-level deployments, the model is available across several cloud and AI infrastructure platforms, including Baseten, Crusoe Cloud, DeepInfra, Fireworks, Nebius, and Together AI. These providers offer a hosted path for teams looking to bypass the complexities of managing underlying infrastructure while still maintaining the benefits of the tuned harness.
The shift toward these specialized, open-stack agents reflects a maturation in enterprise AI. As the technology moves from experimental phases into core business processes, the emphasis is increasingly placed on reliability, cost-efficiency, and the ability to audit AI decision-making. By providing a blueprint that allows for continuous evaluation and adjustment, NVIDIA and LangChain are positioning their respective technologies to meet the demand for agents capable of high-stakes, autonomous execution. Developers interested in building or migrating their agentic AI workflows can find the latest documentation and reference blueprints through the official LangChain and NVIDIA developer portals.
Keep reading