LangChain and NVIDIA have released a new blueprint for building AI agents that costs roughly 90 percent less to run than competing systems while maintaining strong performance. The NemoClaw for LangChain Deep Agents blueprint combines three key components: LangChain's agent code, NVIDIA's Nemotron 3 Ultra model, and NVIDIA's OpenShell runtime environment. In testing, this combination achieved a performance score of 0.86 at a cost of $4.48 per task, compared to $43.48 for the next best-performing alternative.

What Problem Does This Blueprint Solve for Enterprises?

As companies move AI agents from experimental pilots into production systems, they face a critical challenge: the infrastructure around the model becomes as important as the model itself. Agent memory systems, workflow designs, evaluation datasets, and runtime policies all represent proprietary knowledge that shapes how a company competes. In closed ecosystems controlled by cloud providers, teams lose control over this intellectual property and cannot fully customize it for their specific business needs.

The NemoClaw blueprint addresses this by giving enterprises ownership of the entire agent stack. Rather than relying solely on expensive cloud APIs, teams can run agents on open-source models they control, tune the system around their specific workloads, and apply governance policies from day one. This shift matters because it transforms AI agent costs from a fixed, unpredictable cloud bill into something teams can actively manage and optimize.

How Does Lower Cost Change the Way Teams Build Agents?

When inference costs are high, teams make fewer experiments and take fewer risks. They run smaller evaluation suites before deployment, compare fewer model variants, and avoid building specialized agents for specific domains because the operating cost becomes prohibitive. Lower costs fundamentally change this calculus.

With the NemoClaw blueprint's cost efficiency, teams can afford to run larger evaluation suites throughout the agent development lifecycle. They can test changes to prompts, tools, and model configurations more frequently. After deployment, they can monitor real-world behavior, make fixes when needed, and create tests to prevent regressions. This continuous improvement cycle becomes economically viable when each iteration costs a fraction of what it did before.

The practical implication is that the best agent system for a given workload emerges from optimizing across multiple dimensions simultaneously: quality, cost, speed, and governance. A real-time customer support agent has very different requirements than a coding agent running background tasks, and lower costs make it practical to build specialized solutions for each use case.

What Components Make Up the Blueprint?

• Model Layer: NVIDIA Nemotron 3 Ultra, an open model that enterprises can run, customize, and optimize for their specific workloads without relying on proprietary cloud services.
• Agent Harness: LangChain Deep Agents Code, which provides the framework for long-running agents including planning, tool use, memory management, and task execution, with a profile specifically tuned for Nemotron 3 Ultra.
• Governed Runtime: NVIDIA OpenShell, a secure runtime environment that enforces policy-based rules for how agents interact with tools, systems, and data, with complete logging for auditability and compliance.

Together, these layers create a path for enterprises to build, deploy, measure, govern, and continuously improve agents in production environments. The blueprint is supported by an ecosystem of partners including Baseten, Fireworks, Nebius, Crusoe, DeepInfra, and Together AI, who help enterprises serve these models at scale and adapt the blueprint for business-critical applications.

Why Are Enterprises Concerned About Agent Governance?

Regulated industries face particular pressure to move agents from isolated pilots into production while maintaining strict control over how those systems operate. Companies in finance, healthcare, and other compliance-heavy sectors need transparency into agent decision-making, control over where data and inference run, and the ability to prove to regulators or boards that the system operates within defined boundaries.

"EY clients in regulated industries are ready to move agentic AI out of isolated pilots and into production and are often constrained by governance, security, and the ability to prove control to a regulator or a board. Open agent architectures matter because they give enterprises transparency into how agents operate, control over where data and inference run, and the freedom to deploy on their own terms without committing to a closed stack," said Geoff Vickrey, Global Chief Commercial Officer at NVIDIA and EY.

Geoff Vickrey, Global Chief Commercial Officer, NVIDIA and EY

The NemoClaw blueprint addresses these concerns by building governance into the runtime from the start. OpenShell enforces policy-based privacy and security rules and logs each agent action, creating an auditable record of how the system behaves. This approach allows enterprises to meet regulatory requirements while maintaining the flexibility to deploy agents wherever they operate.

How Does This Compare to Running Agents on Local Hardware?

A complementary development shows how enterprises are applying similar thinking to on-premises infrastructure. Dell's Deskside Agentic AI system, launched in May, lets work groups run production-ready agents on Dell workstations using the same NemoClaw open-source stack. This approach targets teams that cannot send work to public cloud services, such as engineers running coding agents that must keep source code in-house or researchers analyzing sensitive data subject to privacy regulations.

Analysis by Signal65 and Futurum found that running agents locally on this infrastructure can reduce token costs by up to 87 percent over two years compared to public cloud APIs, with break-even occurring in as little as three months. The key insight is that once a prototype moves from a workstation to larger infrastructure, it can migrate to Dell PowerEdge servers in the data center without requiring architectural changes, because both environments use the same OpenShell governance layer.

This flexibility matters because it lets enterprises make cost-conscious decisions about where each workload runs. Everyday tasks execute on hardware the company already owns, where the marginal cost of a token approaches the cost of electricity. Larger frontier models are called in like specialists for specific high-value tasks, rather than serving as the default workhorse for all agent operations.

How to Optimize Agent Costs Across Your Organization

• Start Local: Begin by running validated workflows for coding, research, and private assistants on compact, efficient models with 30 billion parameters on local hardware, reserving larger models for tasks that genuinely require their capabilities.
• Govern Early: Implement policy-based governance and logging from the first deployment, not as an afterthought, so that compliance and auditability are built into the system architecture rather than bolted on later.
• Scale Smart: Design your token strategy before the first agent ships into production, making deliberate decisions about which workloads run where and what governance applies to each, so that costs become something you manage rather than something that happens to you.

These principles reflect a broader shift in how enterprises think about AI infrastructure. Token consumption is no longer just an IT concern; it has become a board-level financial question. Companies that make deliberate decisions about where agents run and what policies govern them can control costs. Those that default to cloud APIs for all workloads face unpredictable bills that scale with agent complexity and volume.

"Super agents have arrived. With an open model like NVIDIA Nemotron, a LangChain harness, the NVIDIA OpenShell runtime, and a company's own data, every enterprise can build custom agents that understand its business, use its tools, and turn knowledge into action. The future of AI won't be one-size-fits-all. Companies will use AI cloud services and build their own AI, shaped by their proprietary data, know-how, and workflows, and run it safely and securely wherever they operate," said Jensen Huang, Founder and CEO of NVIDIA.

Jensen Huang, Founder and CEO of NVIDIA

The NemoClaw blueprint and the ecosystem of partners supporting it represent a significant shift in how enterprises approach agent deployment. By combining an open model, a tuned agent framework, and a governed runtime, teams gain the ability to own their agent infrastructure, optimize it for their specific needs, and control costs from the start. For companies building agents at scale, this approach offers a path to production that balances performance, cost, governance, and competitive advantage.