Summary: LangChain and NVIDIA released the NemoClaw for LangChain Deep Agents blueprint on July 8, 2026. The stack combines open Nemotron 3 Ultra, the Deep Agents Code harness, and the policy-governed OpenShell runtime. LangChain’s eval reported 0.86 aggregate for $4.48 versus $43.48 for the nearest competing model.
The “NemoClaw for LangChain Deep Agents” blueprint landed on July 8, 2026, a joint release from LangChain and NVIDIA pitched at enterprises that want agent systems both open and governed. This is no ordinary framework update: rather than a single tool, the package is a three-layer stack — model, agent harness, execution environment — and the two headaches enterprises feel most, governance and cost, are baked straight into the architecture. NVIDIA CEO Jensen Huang kept the framing blunt: “Super agents have arrived.”
Inside the Three Layers: Open Model, Agent Harness, Governed Sandbox
Nemotron 3 Ultra, an open model, forms the base layer, putting deployment, fine-tuning, and optimization into enterprise hands. LangChain’s Deep Agents Code (“dcode”) occupies the middle: an open-source harness that carries planning, tool use, memory, and long-running task execution, shipping alongside a harness profile tuned for this model — what the companion post calls “tune the harness, not the model.” NVIDIA OpenShell completes the stack as a policy-governed sandboxed runtime that sets the terms for how agents reach tools, systems, and data. The full blueprint sits on NVIDIA’s build platform, runs out of the box, and every layer can be audited, adjusted, or swapped independently.
A Tenfold Cost Gap, Straight From the Evals
LangChain’s published numbers pull no punches: on its own agent eval suite, the open model running with the tuned Deep Agents harness scored 0.86 aggregate for $4.48, while the nearest competing model rang up $43.48 — close to a tenfold spread. In agentic settings, inference cost calls the shots: it sets how large an eval suite can grow, how many variants can go head-to-head, and whether long-running tasks ever make commercial sense. Lin Qiao of Fireworks says in the launch post that inference speed and cost “directly determine whether an agent is viable in production.” Once that cost falls away, larger eval matrices and domain-specific agents come within reach.
dcode’s Approach to Sensitive Code
The companion post lays out the governance mechanics. One command, nemo-deepagents onboard, raises the sandbox, installs dcode, and hooks the model in through NemoClaw’s managed inference. The governance rules read: networking denies everything by default and each egress request needs individual approval; sensitive actions sit behind human approval gates; any run can be snapshotted into NemoClaw’s per-session audit logs; credentials stay with NemoClaw outside the sandbox, where “the agent itself never touches them.” A custom Dockerfile can build the sandbox, covering Java, COBOL, Maven, Gradle, and further enterprise stacks, and local repositories are pushed in by command — code remains inside a boundary you control, while data residency and infrastructure stay yours.
Legacy Modernization Takes the Flagship Spot
Legacy modernization holds the blueprint’s flagship scenario, with COBOL-to-Java as the canonical case. Work moves in waves: assess and document first, decompose along business domains, refactor step by step, and send every diff through human review and approval before the following wave starts. Also on the list: .NET and Windows-to-Linux migrations, cloud re-platforming, dependency and end-of-life framework upgrades, test repair and coverage work, security patching that leaves audit-friendly records, and governed internal coding assistants. Setup is deliberately uneventful: a wizard presents NVIDIA Endpoints or any OpenAI-compatible endpoint, together with the default OpenShell resource profile and a “Balanced” policy tier. For enterprises uneasy about letting agents near core systems, the “a human approves every wave” cadence is comfort engineered on purpose.
The Ecosystem, and the Case for Openness
The ecosystem arrived together with the blueprint: EY is building out an implementation practice, and Baseten, Fireworks, Nebius, Crusoe, DeepInfra, and Together AI carry production serving of the model. EY’s Geoff Vickrey makes the argument that open architectures give enterprises “transparency into how agents operate.” Openness is the core pitch — fine-tune the model weights, replace the harness, run the runtime yourself, and no single layer on its own locks you in. The more revealing demonstration is governance recast from 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 leaning on 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 weighing agent platforms can pick up this blueprint and use it as a ready-made checklist.