Jensen Huang: Why companies need open agent systems

First, before I answer, I want to congratulate you for all the work you do. If you look at the last six months, we could both agree that although we've been working in AI for 15 years, the last six months changed everything. All the technology, large language model advances, scaling, breakthroughs, omni models, multimodality – it was fantastic. But the last six months is where everything came together, and now finally AI is useful. When AI is useful, every company and enterprise wants to get their hands on it. Now the question is how, and this is where LangChain comes in. You always had a vision that the large language model was the essential ingredient, but to turn it into a useful product you have to surround it with a harness. We used LangChain to turn a large language model into a promptable API, build RAGs, and step by step that led to today's agents. The big breakthrough in the last six months are agentic systems that are grounded on information and knowledge, can use tools to search, have memory, safeguards, and iterate until the job is done. But it needed models that have reached a level of capability where everything comes together into a flashpoint. Claude Code really brought the imagination of agentic systems. OpenClaw was a big deal, and all the work you did with Deep Agents – we use that ourselves – all came together. The reason we dedicate ourselves to open systems is because AI is a fundamental technology; it can only be useful if applied in many different use cases. We imagine a world where scientists, digital biologists, designers, roboticists, students, researchers, and enterprise IT can all use agentic AIs to solve domain-specific problems. We either embed specialized domain knowledge that is not available outside, or we put AI into a flywheel where we use it, it gets smarter, becomes more useful. We use it more, it gets even smarter – learns over time. We imagine a future where AI has a foundation, and the work from Anthropic, OpenAI, Google is fantastic, but there are also specialized, domain-specific, proprietary AIs that people want to build, and we want to enable that world.

Specialization starts with intelligence that is good enough. That's why we worked on Nemotron and love that you are part of the founding team of the Nemotron Coalition. We made Nemotron Ultra, which is a great model as a start, but it becomes incredible when you put the LangChain framework around it so you ground it on domain-specific information. An intelligent person becomes super useful when given access to important information. So access to information is important. Putting it into a flywheel, maybe post-training the model inside the LangChain harness so it becomes good at applying the harness for that task. I think this moment has arrived, but we need an open harnessing system we can build ourselves and improve over time.

Yeah.

Ooh.

I know. I am so proud. But it is so incredible. Thank you.

The benefit of cost comes in a couple of ways. When you have cost-effective intelligence, people use more of it. When you have a cost-effective agent, you can iterate across a larger search space, and the answer can actually be better. Nemotron is cost-effective because it is so fast and computationally efficient. When it's computationally efficient, it can explore larger spaces – no different than when somebody can think fast, they can explore more space and find a better answer. That's the incredible benefit of Nemotron 3 Ultra inside the LangChain framework and harness inside Deep Agents. It can think so quickly, explore so quickly, iterate efficiently, and find better answers. I'm really excited we created a model near the frontier, but adapting the environment around Nemotron made it deliver frontier capabilities. For humans, it's the same: we hire the smartest people, but we give them access to tools and information, and create conditions for them to achieve their full potential. You adjust the environment, not just the model. That's where LangChain came in.

The frontier models are getting better all the time, and I fully expect them to be unbelievably good. They still have a long runway of improving models with scaling laws, harnesses, memory technology, compaction, retrieval augmented generation, knowledge graphs. There are still many incredible advances being implemented into frontier model APIs. The way I think about it is I always start all my work with the frontier because it's useful. I know what's potential; it costs a bit more, but my time to getting work done is fast. Over time, I want to add sub-agents that are super agents at certain skills. We have optimization problems in our company related to supply chain, chip design, floor planning. These are insanely hard, so we create super sub-agents with LangChain Deep Agents and Nemotron 3 inside, connected to specialized tools. That super agent is built for one job – not booking travel, but optimizing supply chain. I need LangChain and Nemotron 3 Ultra with proprietary knowledge and skills. A whole team refines that. That defines a company – a collection of super proprietary workflows. Now we can have LangChain with Deep Agent and Nemotron 3 inside, giving all the control and efficient access to tools. That's the future.

As soon as it gets good enough. Start with Claude Code and Codex and use it as long as you can. For many things you never have to replace because they keep getting better. In the future, like today, companies have employees for domain specialization, but also hire consultants, license external tools, outsource work. That will be the future for AI. Will we continue to use frontier models? Absolutely, tons of it. But will we also create specialized super agents with LangChain and Nemotron 3 Ultra that could be your crown jewels? Absolutely true.

Yeah. In the future most companies will be built on harnesses. LangChain will become the tool that creates the operating system for the company, and everyone will use LangChain to create specialized harnesses representing workflows. That harness becomes autonomous, agentic, and much more efficient.

That's right.

This is incredible, this is the big breakthrough. What you just described is a future where once the harness is built, doing the work and part of the business process, the question is how to get even better. You can improve the information, tune the harness, but now you can also improve the AI model inside the harness. I think that's a complete breakthrough – a capability that never existed before. It will take enterprise business processes and start tuning the flywheel.

Every company is built on domain-specific specialized intellectual property. Intellectual property is intelligence. Every company is built on a foundation of specialized intelligence. You can't not control and improve it. Outsourcing that intelligence makes no sense. There is general intelligence for general things – coding, writing – and we apply those foundational skills for specialized domain intelligence. That's where LangChain and Nemotron come in. Society will have foundational models that are general and available in the cloud, but on top of that we must build our own specialized capabilities with open tools. You can't outsource enhancing your intelligence. That future is not one or the other; it's complementary. We are making sure automated intelligence is integrated into everything, and we'll all be better.

This is such a huge deal. Hopefully it makes it much easier for enterprises to get up and running. All the key ingredients for building your personal domain-specific super agent – technologies, components, tooling, harnessing, and the blueprint – all put together for you.

Without solving security and access control, it's impossible to deploy. It's no different than hiring a new employee – you onboard them, give them access. We give employees access to tools, networks, information, and connect them to other agents. We provide a skills file, a mission, and previous work. In a lot of ways we are creating an HR system for AI that allows IT and business units to build, improve, and deploy agents inside companies.

It's electrons, not atoms. No consciousness, not awake. It's a tool like a vacuum cleaner roaming the house, doing what I used to do. We have autonomous lawnmowers, dishwashers – we call them dishwasher, a bit like a human. When I first worked, I was a dishwasher. We'll get used to it. Right now we tend to imbue too many human properties. It's software. We know how it works because we create the harnesses. If we didn't understand how something works, how do we make it better every time? We understand these things. The more AI we use, the more people we hire because agentic systems are new skills. Now software engineers build agents instead of coding – they prefer it. Coding is like typing; they become systems engineers, creating evals, benchmarks, guardrails. That's creating a lot of jobs, and my engineers love it.

That's right.

Whether you're a doctor, designer, or software engineer, you are creating an agent. You take mundane work and get the agent to do it, but we are all trying to elevate agents to do things with us that we couldn't do before. That requires imagination, creativity, and technology.

A hundred percent. Ambition, agency. Ambition helps. Yeah.

Today we are announcing a very big deal. We are providing the basic building blocks, all the key ingredients to build super agents. These super agents are domain-specific, belong to you. You build them, improve them, refine them over time. Give them access to proprietary information, and they will do things you can't imagine. We've created a world-class language model, a framework called LangChain Deep Agents fine-tuned to expose the full potential of Nemotron 3 Ultra, a blueprint to help everyone do that, and the OpenShell runtime that keeps it secure with acceleration stacks integrated. Every company and every developer can now create these super agents and deploy them anywhere – cloud, on-prem, on DGX Spark, DGX Station, or your own supercomputer. All the pieces are here. There are no excuses not to engage.

Thank you. Good job. Proud of you guys.