NVIDIA & LangChain launch open stack for AI agents

NVIDIA and LangChain have launched NemoClaw for LangChain Deep Agents and a tuned profile for NVIDIA Nemotron 3 Ultra, giving businesses an open stack for building AI agents.

LangChain tuned its Deep Agents harness for Nemotron 3 Ultra rather than retraining the model itself. The work involved changes to system prompts, tool descriptions and middleware after analysis of execution traces from its public Deep Agents benchmark.

According to the companies, the tuning delivered the highest accuracy among open models on LangChain's Deep Agents benchmark. They added that the model reached business-task parity with the highest-scoring closed models while cutting inference cost per run to one-tenth of leading closed alternatives.

The announcement reflects a familiar debate in enterprise AI: whether companies should rely on closed systems from model providers or assemble more of the software stack themselves. NVIDIA and LangChain are making the case for the latter by combining an open model, an open orchestration layer and an open runtime that can be deployed across environments.

Harness tuning

LangChain described the project as an effort to improve the system around the model rather than alter the model weights. Its team ran Nemotron 3 Ultra through the Deep Agents benchmark, identified where performance dropped and adjusted the surrounding harness to improve results.

Developers already using LangChain Deep Agents can adopt the tuned profile directly without retraining the model. NVIDIA said the broader NemoClaw blueprint packages that work for enterprises building specialised agents for their own workflows.

NemoClaw for LangChain Deep Agents combines LangChain Deep Agents Code, tuned for Nemotron 3 Ultra, with NVIDIA OpenShell, described as a secure runtime for executing agent actions. The stack is aimed at businesses that want to customise, govern and run agent systems on their own infrastructure or in their preferred cloud environments.

The distinction is becoming more important as AI systems move beyond responding to prompts and toward carrying out actions inside business software. In that setting, control over memory, tool use, evaluation and runtime behaviour can matter as much as the underlying foundation model.

LangChain said its platform records more than 200 million monthly downloads, giving it a large developer base for agent orchestration. That reach could help the tuned Nemotron profile spread quickly among teams already testing AI agents in internal operations and customer-facing products.

Early adopters

Several companies are already using specialised agents in their systems. NVIDIA said Abridge, Amdocs and Box are embedding them directly into their platforms, while EY is expanding its implementation work around NemoClaw blueprints for LangChain Deep Agents.

EY's role highlights the services opportunity emerging around agent deployment. Large organisations typically need help adapting these systems to industry workflows, setting controls around their use and evaluating outputs before they are used in higher-stakes settings.

Hosted access to Nemotron 3 Ultra is also available through inference providers including Baseten, Crusoe Cloud, DeepInfra, Fireworks, Nebius and Together AI. That gives developers a direct route to running the tuned harness in production without managing all the infrastructure themselves.

For NVIDIA, the release extends its push to make its AI software stack more central to enterprise deployments rather than limiting its role to chips and model hosting. For LangChain, it offers a benchmark-based case for tuning agent infrastructure around a model to improve results and lower operating costs.

Harrison Chase, Co-Founder and Chief Executive Officer of LangChain, said the work supports the view that agent performance depends on more than the model alone. "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."