NVIDIA and LangChain Ship NemoClaw: An Open Stack for Deep Agents
Source: NVIDIA Blog / LangChain Blog — published July 8, 2026
Summary
NVIDIA and LangChain jointly launched NemoClaw, an open-source reference stack for running long-lived agents securely, plus a "Deep Agents Blueprint" that pairs it with NVIDIA Nemotron 3 Ultra and LangChain's Deep Agents Code (dcode) harness. On LangChain's agent eval suite, Nemotron 3 Ultra tuned with the Deep Agents harness scored 0.86 aggregate at $4.48 per run, versus $43.48 for the next-closest competitive model — roughly 10x lower inference cost while matching accuracy of leading closed models. The pitch is an integrated, tunable stack (model + harness + sandboxed runtime) rather than a single benchmark win.
Key Takeaways
- Three-layer stack: Nemotron 3 Ultra (open model layer), LangChain Deep Agents Code (harness layer — planning, tool use, memory, task execution), and NVIDIA OpenShell (secure sandboxed runtime with tool/data access policy) — the claim is that tuning all three together, not just swapping models, is what drives the cost/performance gap.
- NemoClaw itself is runtime-agnostic: the GitHub repo describes it as a way to run agents "like Hermes, LangChain Deep Agents, and OpenClaw" more securely inside OpenShell sandboxes, with a single CLI for onboarding, routed inference, network policy, and lifecycle management — not locked to one harness.
- 10x cost gap is the headline number: $4.48 vs $43.48 per run on LangChain's eval suite, at comparable accuracy to closed frontier models — a concrete data point for teams evaluating open vs. closed models for production agent workloads.
- Positioning against closed-model agent stacks: this is explicitly framed as making an open, self-hostable stack competitive with proprietary agent platforms on both cost and throughput, not just raw model quality.