Jensen Huang: Why companies need open agent systems

NVIDIA founder and CEO Jensen Huang sits down with Harrison Chase to discuss why the last six months finally made AI useful, and what it takes to turn a large language model into a real, deployable product. The path there, Jensen says, is building your own “super agents”: domain-specific systems wrapped in an open harness, grounded in your data, and improved over time. NVIDIA and LangChain also announce a new blueprint for running Deep Agents with Nemotron 3 Ultra inside OpenShell, a secure, open runtime, giving every enterprise the building blocks to create and deploy super agents anywhere.

TL;DR

A fireside interview (video 6 of a same-day-origin 7-video batch) between NVIDIA CEO Jensen Huang and Harrison Chase (LangChain co-founder/CEO), published on the LangChain YouTube channel, 8 Jul 2026. Huang’s central claim: “today, most companies are built on business processes. In the future, most companies will be built on harnesses.” The interview doubles as the launch of a joint NVIDIA + LangChain blueprint — running Deep Agents with Nemotron 3 Ultra inside OpenShell, a secure open runtime — giving enterprises the building blocks (model, harness, blueprint, runtime) to build domain-specific “super agents” anywhere: cloud, on-prem, or on a DGX Spark next to a laptop.

Six substantive threads:

1. Why NVIDIA invests in an open agent ecosystem. Huang frames the last six months as the point where “everything came together” for agentic AI — foundation-model advances plus harness engineering (naming Claude Code and OpenClaw explicitly) plus LangChain’s own build-up from promptable-API wrapper to RAG tooling to agents. NVIDIA’s stated motive for openness: AI is “a fundamental technology” useful only when applied to many domain-specific use cases NVIDIA cannot build itself — scientists, digital biologists, roboticists, enterprise IT all need to build their own specialized, proprietary AIs on an open foundation.
2. Specialization: model, harness, and post-training together. Nemotron Ultra needs “intelligence that’s good enough” first, but becomes useful only inside the LangChain harness, grounded on domain-specific information — and, looking forward, post-trained inside the harness so the model “becomes good at applying the harness around it.” On an internal Deep Agents benchmark, Nemotron 3 Ultra scores 86% vs. Claude Opus’s 87%, at roughly 10x lower cost — DeepSeek and a Minimax model trail at 82-83%. Cheaper, faster inference lets an agent “iterate across a larger search space” and “find better answers,” independent of raw capability.
3. When to specialize: start with the frontier, then build super sub-agents. Huang’s own practice: start every task with a frontier model (Claude Code, Codex) “for as long as I can,” then, once a domain (e.g. supply-chain or chip-design optimization) proves too hard for a general agent, build narrow super sub-agents on LangChain Deep Agents + Nemotron 3, connected to proprietary knowledge and tools. “That thing is built for one job… now I think that defines a company. A company is really about a collection of a whole bunch of these super proprietary, super important workflows.”
4. Companies built on harnesses, not business processes. The interview’s headline claim: LangChain becomes “the tool that creates the operating system for the company,” and the harness inside a workflow “becomes autonomous, agentic, much more efficient.” Every company’s specialized intellectual property is its intelligence — “you can’t possibly not continue to control it, improve it, make it better,” and outsourcing that intelligence “makes no sense,” whether for a person, company, or country. General skills (coding, writing) are foundation-model territory; the proprietary, specialized layer on top requires open tools you own.
5. Runtime, security, and access control. Huang frames the runtime as the unglamorous but load-bearing final layer — “without solving the security, the access control, it’s impossible to deploy,” drawing an explicit parallel to employee onboarding (access to files, networks, tools scoped by role). This is presented as the reason OpenShell exists: a secure sandbox IT organizations can control. In the same breath Huang uses HR-system language (“we are creating an HR system, if you will, for AI… a skills file”) while, minutes later, explicitly rejecting deeper anthropomorphizing.
6. How much to anthropomorphize agents, and why more AI means more jobs. Huang’s position: “it’s electrons, not atoms… it’s not biological, has no consciousness… it’s a tool, like my vacuum cleaner.” He predicts the discomfort fades the way “dishwasher” did. On employment: “the more AI we use, somehow the more people we have to hire” — software engineers now build and orchestrate agents (evals, benchmarks, guardrails) rather than write Python, work Huang says his engineers prefer. Chase adds the sharper framing: most current agent usage still automates what people did before; the larger unlock is doing what “we couldn’t do before.”

What was actually ingested

Full ASR transcript (auto-generated captions), ~26:35 runtime, 15 chapters. A manual English caption track also exists per the raw file’s caption_tracks: but was not the track fetched; treat transcript wording as ASR-cleaned, not verbatim-manual.

Linked entities and concepts

• Channel/publisher: LangChain (8th source).
• Interviewer: Harrison Chase (4th source — LangChain co-founder/CEO).
• Created NVIDIA entity — first substantive source (central subject: NVIDIA’s CEO, products, and joint blueprint announcement), following the same first-appearance-as-central-subject precedent as Khan Academy and PwC. NVIDIA had previously appeared only in passing, as silicon named alongside AMD, in Hoffman.
• Dangling (single-source mention, deferred, per the person-entity second-source rule — central-subject status does not itself trigger promotion for individuals, precedent: Sal Khan on the Khan Academy source): Jensen Huang (NVIDIA founder/CEO — central subject of this interview, but no prior wiki source; will promote on a second substantive Huang-authored or Huang-centric source).
• Concepts: agent-harness, enterprise-ai-adoption, ai-employment-effects, dynamic-capabilities.
• Products/terms new to the wiki, held in body prose rather than promoted to entities: Nemotron 3 Ultra (NVIDIA open-weight model), OpenShell (secure open agent runtime), the Deep Agents + OpenShell blueprint (joint NVIDIA/LangChain announcement), DGX Spark / DGX station (NVIDIA hardware named as deployment targets).

Relationships

See frontmatter. Six typed supports edges: two to existing Harrison Chase / LangChain sources (ADLC, Interrupt 26) on the shared harness/model/context vocabulary; one to Hoffman on independent vendor-CEO convergence around frontier-vs-specialized model choice; one to BCG on the anthropomorphizing-AI debate; one to AWS NYC on runtime security/access-control as an enterprise-deployment precondition; one to the Claude Platform team’s own agentic-infrastructure panel on the shared harness-thinning mechanism and MCP-as-agent-interface, published two days after this interview. Considered but not linked: LangChain Interrupt (same conference-adjacent LangChain-channel cluster and shares an open-weight-models tag, but no specific shared claim, quote, or data point beyond generic thematic proximity — too thin for a typed edge).