Written by: Techub News Compilation

Introduction

Recently, NVIDIA founder and CEO Jensen Huang had an in-depth conversation with the founder of the AI application framework LangChain. This dialogue occurred at a critical juncture in the rapid development of AI agent technology. Over the past six months, with breakthrough advancements such as Claude Code and OpenClaw, agent systems are transitioning from concept to large-scale enterprise applications. As a global leader in AI computing power and ecosystem, Huang's voice elucidated why and how NVIDIA is investing in an open agent ecosystem and announced a key blueprint aimed at lowering the barriers for enterprise applications. His insights hold significant strategic guidance for any company looking to deeply integrate AI into its business processes.

Summary

• The past six months have been a turning point for AI, moving from technological breakthroughs to widespread applicability, with an explosion in enterprise application demand.
• The core competitiveness of enterprises lies in their proprietary knowledge and processes, hence they must build and control their own “super agents” based on open technologies.
• The future of enterprises will be built on “agent frameworks” rather than traditional business processes; an open stack is the cornerstone for ensuring control, security, and continuous evolution.
• NVIDIA is building a complete open agent stack by providing world-class models, frameworks, blueprints, and secure runtimes, empowering enterprises to create proprietary agents.

The Moment AI Becomes Truly Useful: The Leap from Models to Agent Systems

Jensen Huang opened the discussion by pointing out that although AI technology has developed for fifteen years, the past six months have changed everything. Advances in large-scale language models, multimodal capabilities, and scalability breakthroughs have converged at this moment, making AI finally "useful". When AI becomes useful, every business around the globe desires to harness it.

However, the question remains, “how to achieve this?” Huang believes that pure large language models are just raw materials; to transform them into useful products, they must be surrounded by a "framework". He recalled the collaboration with LangChain: initially transforming large models into promptable APIs, then building RAG (retrieval-augmented generation) systems, and developing to today's agents step by step. The real breakthrough is the emergence of agent systems that are information- and knowledge-based, capable of using tools for searches, managing memory, ensuring safety, and iterating to complete tasks.

All of this requires model capabilities to reach a critical point, and the emergence of technologies like Claude Code has ignited the imagination around agent systems. Huang stressed that NVIDIA has long been committed to building open systems because AI is a foundational technology that can only deliver maximum value when applied across vast and varied use cases. His envisioned future is one where scientists, digital biologists, designers, robotics experts, students, researchers, and enterprise IT can use agents to solve problems in specific domains.

The Path to Specialization: The Fusion of Models, Frameworks, and Proprietary Knowledge

When asked about the best way to achieve system specialization, Huang pointed out that specialization starts with sufficiently good foundational intelligence. This is precisely why NVIDIA launched the Nemotron model series. He cited Nemotron-3 Ultra as an example, explaining that an excellent model is the starting point, but when placed within the LangChain framework and "brought down to earth" based on domain-specific information, it can become incredible.

“A smart person becomes super useful when we give them access to especially important information,” Huang analogized. Hence, information access is crucial. Furthermore, models can be placed within a “flywheel”: conducting subsequent training on the model within the LangChain framework to enhance its proficiency in completing specific tasks using the surrounding framework. This means that enterprises not only can optimize the framework (prompts, tools) but also can optimize the model itself in the future, representing a complete breakthrough.

The founder of LangChain added their practice: by adjusting the framework to accommodate different models (since different models require different prompts and tools), they successfully achieved an 86% score for Nemotron-3 Ultra in internal benchmarking, very close to Claude Opus’s 87%, yet at just a tenth of the cost. This highlights the excellent balance struck by open-weight models between performance and cost.

Huang delved into the multiple manifestations of cost advantages: when agent costs are efficient, people will use them more frequently; when agents are cost-effective, they can iterate in a larger search space, potentially finding better answers. Nemotron’s speed and computational efficiency make it cost-effective, allowing for rapid exploration and iteration to find optimal solutions. This is akin to a quick-thinking person being able to explore more possibilities.

Leading Models and Proprietary Agents: Complementary Enterprise AI Strategies

In response to the question of “whether to always use open-source models,” Huang provided a clear dual-track strategy. He affirmed that leading models will continue to improve, with scaling laws remaining effective, and technologies such as frameworks, memory handling, RAG, knowledge graphs, etc., evolving rapidly.

His personal workflow begins with always starting from the use of leading models. Leading models are useful, quickly presenting potential, though at a slightly higher cost, they can significantly reduce the time taken to complete work. However, over time, he finds the need to add “sub-agents”—agents that become “super specialists” in certain skills.

He cited extremely complex issues like supply chain optimization and chip design layout optimization within NVIDIA as examples, explaining that it is impossible to rely on a generic AI to brute-force solve them. Instead, they create these super sub-agents using Deep Agents (based on LangChain), integrated with Nemotron-3, and connect them to specialized tools. This agent is born for one job alone, not responsible for booking travel, focusing solely on optimizing the supply chain.

“This defines a company,” Huang stated, “a company is a collection of numerous super proprietary, super important workflows.” Now, they can use LangChain Deep Agents with built-in Nemotron-3 Ultra to build these workflows, granting teams all the control they need and the ability to efficiently access powerful tools. “This is the future.”

Regarding advice for enterprises, Huang believes that once agents are sufficiently good, they should start specializing. He recommends starting with cutting-edge models like Claude Code and Codex and using them for as long as possible, as they are rapidly advancing. In the future, enterprises will continue to make extensive use of leading models, just like today when outsourcing consultants, authorizing external tools, and subcontracting work. At the same time, enterprises will also use LangChain and Nemotron-3 Ultra to create specialized super agents, which may become the “crown jewels” of the business. Even “consultative” AIs need to be integrated into organizations, with access to context, tools, and data.

Open Ecosystem: The Cornerstone of Enterprises Controlling Their Own Intelligence

Huang emphasized that an open stack is crucial for companies to deeply apply AI. Every company is built on specific domain or specialized intellectual property. “We call it ‘intellectual property,’ and ‘knowledge’ is intelligence.” The core competitiveness of every company is its unique intelligence foundation. Outsourcing this core intelligence is meaningless for individuals, companies, or countries.

The foundation of society will have these foundational models, which will be universal and available on the cloud, which is fantastic. But on this platform, enterprises must build their specialized capabilities, which requires open tools. “I cannot imagine calling a third party when I need to enhance my intelligence. I need to enhance it internally within the company.” Therefore, the future is not a choice between two options, but a completely complementary vision: the coexistence of general foundational models and proprietary agents.

To lower the barriers for enterprise applications, NVIDIA and LangChain announced a blueprint including Deep Agents and OpenShell, as part of the NemoClaw blueprint. This enables enterprises to run Deep Agents with built-in Nemotron-3 Ultra within OpenShell (a secure, open runtime environment). Huang called this a “big deal,” incorporating all the key technologies, components, tools, and frameworks needed to build personalized, domain-specific, proprietary super agents.

From Tools to Organization: Key Considerations for Deploying Agents

Building agent systems involves numerous complex components: large language models, tools, knowledge graphs, memory systems, protection systems, fine-tuning systems, as well as technologies for subsequent training on frameworks, and the framework itself combined with the runtime environment. Huang particularly emphasized the importance of runtime and security.

“Without addressing security and access control issues, deployment is impossible.” He likened this to the impossibility of hiring a new employee without onboarding them or granting access. Companies do not give every employee access to all files and networks. Similarly, for agents, it is essential to grant access privileges to tools, network partitions, information, and connections to other agents or colleagues based on their responsibilities and provide a “skill file” outlining their mission and historical working methods.

In this sense, they are creating a kind of "human resources system" for AI, enabling IT departments and various business units within the company to build, improve, and deploy these agents. Huang also warned that although people often anthropomorphize agents, it is crucial to recognize: they are electronic, not atomic; they have no biology, no consciousness. They are a tool, much like a robotic vacuum cleaner or a dishwasher. Ultimately, we will become accustomed to them, but right now, we are attributing too many human qualities to them.

He optimistically pointed out that more applications of AI actually create more jobs. His software engineers are now more keen on building agents rather than writing Python code. “Coding is like typing; they will reduce typing and become more engineers building these super cool autonomous systems.” They create evaluation systems, benchmark tests, and safety barriers; the very process of bringing AI into the world creates numerous jobs.

Finally, Huang concluded that the blueprint announced today represents a “very big event.” NVIDIA provides all the essential components and key elements required to build super agents: world-class language models, the LangChain Deep Agents framework fine-tuned to fully unleash the potential of Nemotron-3 Ultra, blueprints that help everyone realize their goals, and secure OpenShell runtimes with integrated acceleration stacks.

“Now, every developer in every company worldwide should be able to create these super agents and deploy them anywhere—cloud or on-premises.” Huang stated passionately, asserting that all components are in place with no reason not to dive in. This marks the complete opening of the era for enterprises to build and control proprietary AI agents.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。