Software "agents" act inside business tools; NVIDIA and LangChain say their open model plus tuned orchestration lands at parity with closed rivals, on a benchmark they themselves run.
Every enterprise now faces the same AI build-versus-buy decision: rent a closed model's reasoning per call, or stand up an open stack the company itself governs. NVIDIA and LangChain are pitching the second path with a new blueprint called NemoClaw that they say makes the open option cost-competitive with the closed one.
"Agents," here, means software that takes actions inside a company's existing tools: customer-relationship apps, code repositories, internal dashboards. The question is whether a model can complete a multi-step task across those systems while a human watches, not whether it can answer a prompt. LangChain's Deep Agents framework is the company's attempt to standardize that workflow, and NemoClaw is the named blueprint for running it on NVIDIA hardware.
The joint pitch rests on what was not retrained. According to the NVIDIA and LangChain announcement, only the harness around the model was tuned: the system prompts, the tool descriptions, and middleware shaped from analysis of execution traces. The foundation model itself, NVIDIA's Nemotron 3 Ultra, was left untouched. Both companies use the phrase "tune the harness, not the model," and NVIDIA's developer blog walks through how the traces were turned into a reusable profile. LangChain's companion post frames that as an enterprise trade-off: no retraining means no GPU cluster, no licensing negotiations over derived weights, and no governance questions about a new model release.
On top of that base, the two companies claim two things. The first is that their configuration posts the highest accuracy of any open-weight model on LangChain's Deep Agents benchmark. The second is that it reaches business-task parity with leading closed alternatives at roughly one-tenth the inference cost per run. LangChain also notes that its tooling passes 200 million downloads a month, a company-provided figure meant to suggest the ecosystem can absorb a new blueprint.
The reproducibility artifacts are public. The langchain-ai/deepagents repository ships the evaluation surface, and the EVAL_CATALOG and MODEL_GROUPS files define which benchmarks cover which models. A reader can rerun the harness profile against the catalog and see what changes; the open label is what makes that portability possible.
The headline numbers, on the other hand, are not independently verified. The cost figure is the companies' own comparison against "leading closed alternatives," on benchmarks the companies run, on orchestration the companies wrote. The trade outlet that first carried the joint release is a same-day reprint of the press wire, and no analyst, customer, or competitor is quoted. An enterprise treating this as a procurement signal would run the open evaluation suite against the workloads the team already pays for, and size the cost claim against its own traffic rather than the companies' example agents.
Shipped together, the three layers form an open enterprise-agent stack. Nemotron 3 Ultra supplies the weights. Deep Agents and the NemoClaw blueprint supply the orchestration. NVIDIA's OpenShell runtime supplies an execution environment that can run on a customer's hardware or its cloud of choice. An enterprise could assemble each layer already; NemoClaw ships them tuned together as a single blueprint with a tuned harness in front.
Closed-model pricing has shaped enterprise agent deployment, because each multi-step task multiplies model calls. If the companies' cost figure holds against an enterprise's own traffic, the build path becomes the cheaper path, and the buy path has to answer for what the closed provider is selling beyond a model. Whether it holds is now a question for each buyer's own evaluation, not the vendors'.