LangChain unveils NemoClaw deep agents

LangChain said on July 8 that it had released the NemoClaw Deep Agents blueprint with NVIDIA, adding a packaged reference architecture for enterprises that want to build and run agent systems without relying on a single closed vendor. The launch ties together LangChain’s Deep Agents framework, NVIDIA’s Nemotron model family and NVIDIA OpenShell, a sandboxed runtime for governed execution. LangChain said the package is meant for teams moving agents from demos into production, where cost controls, deployment policy and customization matter alongside model quality. NVIDIA’s NemoClaw site describes the offer as an open blueprint that enterprises can customize, govern and run in different environments. ### What exactly did LangChain launch? LangChain’s announcement described NemoClaw for LangChain Deep Agents as a blueprint rather than a standalone model or hosted application. The company said it gives enterprises a reference architecture to build, evaluate and deploy advanced open agent systems. NVIDIA’s blueprint card says the package is aimed at “Deep Agents Code,” an open-source terminal coding agent built for software engineering tasks such as planning, editing files, running commands and testing code. (prnewswire.com) NVIDIA’s documentation says NemoClaw adds onboarding, lifecycle management, policy controls and inference routing around LangChain Deep Agents Code running inside OpenShell containers. LangChain’s Deep Agents documentation says the underlying framework is built for long-running, multi-step work with planning, context management, subagents and memory. ### Which pieces of the stack matter most? NVIDIA’s materials say the blueprint packages LangChain Deep Agents, a tuned Nemotron 3 Ultra profile and the OpenShell runtime. (prnewswire.com) LangChain’s provider documentation says the broader partnership spans sandboxed agents with OpenShell, LangGraph acceleration primitives, NeMo Agent Toolkit optimizations and post-training workflows. OpenShell is the governance layer in that package. NVIDIA’s quickstart and overview pages say the runtime is designed to keep source code, credentials, execution and network access under enterprise control, while NemoClaw adds a command-line interface, managed configuration and policy enforcement for supported agents. (docs.nvidia.com) ### Where does the “10x lower inference cost” claim come from? LangChain’s release said the blueprint delivers benchmark-leading performance with more than 10x lower inference costs. (nvidia.com) NVIDIA’s marketing page repeats that claim, and a secondary report citing LangChain’s figures said the system posted a 0.86 agent-eval score at $4.48 per run, about 10 times below the next closest model cost. (build.nvidia.com) LangChain’s recent Deep Agents updates provide some context for that pitch. In a June product post, the company said Deep Agents 0.6 added a code interpreter, harness profiles, streaming changes and ContextHub support aimed at making agents faster, cheaper and more scalable. ### Why does this matter to enterprise buyers? NVIDIA’s NemoClaw overview says the system can run in cloud, on-premises environments, RTX PCs and DGX Spark, and can pair hosted inference providers or local model endpoints with a hardened sandbox and declarative egress policy. (prnewswire.com) Those features address procurement questions that usually sit outside model benchmarks, including where an agent runs, how network access is constrained and whether a team can swap models or keep workloads local. (langchain.com) LangChain’s release framed the blueprint as a response to enterprises building systems around agents rather than buying a single model endpoint. Because the package is described as open and customizable, the launch gives buyers another option when comparing cost transparency, deployment control and vendor dependence across agent platforms. That framing is an inference from the product materials, which emphasize openness, governance and model choice rather than an exclusive managed service. (docs.nvidia.com) ### Where can teams actually try it? NVIDIA’s build portal lists “NemoClaw for LangChain Deep Agents Code” as an available blueprint, and LangChain’s blog published a July 8 post called “Deep Agents Code on NemoClaw: a governed blueprint for your most sensitive code.” LangChain’s integration docs also point developers to Python adapters, setup notes and notebook walkthroughs for OpenShell Deep Agent deployments. NVIDIA’s quickstart pages say teams can install NemoClaw and run a first Deep Agents sandbox with the `nemo-deepagents` tooling. (prnewswire.com) LangChain’s Deep Agents page and docs continue to serve as the entry point for the underlying framework, while NVIDIA hosts the blueprint, runtime guidance and deployment controls. (docs.nvidia.com) (build.nvidia.com)