On July 8, 2026, LangChain and NVIDIA launched the “NemoClaw for LangChain Deep Agents” blueprint, positioned as a way for enterprises to build open, governed agent systems. What separates it from a routine framework update: it is not one tool but a three-layer stack spanning model, agent harness, and execution environment, with the two pains enterprises care about most — governance and cost — designed into the architecture. NVIDIA CEO Jensen Huang kept it blunt: “Super agents have arrived.”
A Three-Layer Stack: Open Model, Harness, Governed Sandbox
The first layer is the open Nemotron 3 Ultra model, which teams can run, customize, and optimize themselves. The second is LangChain’s Deep Agents Code (“dcode”), an open-source agent harness handling planning, tool use, memory, and task execution for long-running agents; the blueprint ships with a harness profile tuned specifically for Nemotron 3 Ultra — a companion post calls the approach “tune the harness, not the model.” The third layer is NVIDIA OpenShell, a policy-governed sandboxed runtime that controls how agents touch tools, systems, and data. The whole blueprint is available on NVIDIA’s build platform, ready to run, and every layer can be audited, tuned, or swapped out.
Evals Show a Tenfold Cost Gap
LangChain’s numbers are aggressive: in its own agent eval suite, Nemotron 3 Ultra with the tuned Deep Agents harness scored 0.86 aggregate at $4.48 in cost. The next closest model cost $43.48 — roughly a tenfold gap. In agentic workloads, inference cost decides how large an eval suite you can afford, how many variants you can compare, and whether long-running tasks make business sense at all. Lin Qiao of Fireworks says in the launch post that inference speed and cost “directly determine whether an agent is viable in production.” Once cost drops, bigger eval matrices and specialized domain agents become practical.
How dcode Handles Sensitive Code
A companion post details the governance mechanics. A single command, nemo-deepagents onboard, builds the sandbox, installs dcode, and connects it to the model through NemoClaw’s managed inference. The governance list: networking is deny-by-default, with egress approved per request; sensitive actions hit human approval gates; every run can be snapshotted into NemoClaw’s per-session audit logs; and credentials are held by NemoClaw outside the sandbox, so “the agent itself never touches them.” Sandboxes can be built from a custom Dockerfile to support Java, COBOL, Maven, Gradle, and other enterprise stacks, while local repositories are uploaded into the sandbox by command — code never leaves a boundary you control, and data residency and infrastructure stay yours.
Legacy Modernization Is the Flagship Use Case
The blueprint’s flagship scenario is legacy modernization, COBOL-to-Java being the canonical example. The workflow advances in waves: assess and document first, decompose by business domain, refactor incrementally, and review and approve each diff before the next wave begins. Other named targets include .NET and Windows-to-Linux migrations, cloud re-platforming, dependency and end-of-life framework upgrades, test repair and coverage work, security patching with audit-friendly records, and governed internal coding assistants. Setup is deliberately uneventful: a wizard offers NVIDIA Endpoints or any OpenAI-compatible endpoint, the default OpenShell resource profile, and a “Balanced” policy tier. For enterprises wary of letting agents near core systems, the “a human approves every wave” cadence is clearly a deliberate comfort feature.
Ecosystem and Openness
The surrounding ecosystem shipped alongside: EY is building an implementation practice, while Baseten, Fireworks, Nebius, Crusoe, DeepInfra, and Together AI handle production serving of the model. EY’s Geoff Vickrey argues open architectures give enterprises “transparency into how agents operate.” Openness is the core pitch — model weights can be fine-tuned, the harness can be replaced, the runtime is yours to run, and no single layer locks you in. The more interesting demonstration, though, is turning governance from an after-the-fact audit into an architectural property: deny-by-default networking, credentials outside the sandbox, and per-session logs are enforced by the runtime rather than left to team discipline. For teams that have watched coding agents wander past their remit, that shift — from trust to enforced boundaries — is the real product. Teams evaluating agent platforms can treat this blueprint as a ready-made checklist.
Sources
- LangChain and NVIDIA launch the NemoClaw Deep Agents Blueprint — LangChain
- Deep Agents Code on NemoClaw — LangChain
- NemoClaw for LangChain Deep Agents Code — NVIDIA Build
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.