Open Agent Systems: Why Companies Need Control Over AI
Jensen Huang explains why enterprises must build specialized AI agents using open ecosystems rather than relying solely on frontier models. The key is combining a capable base model with a customizable harness (LangChain), domain-specific knowledge, and proprietary tools—creating a flywheel where companies continuously improve their own AI systems while maintaining full control over their intellectual property.
The AI Inflection Point: From Useful to Essential
The Last Six Months Changed Everything
After 15 years of AI research, the past six months represent a breakthrough where AI finally became genuinely useful. This triggered universal demand from enterprises worldwide to adopt AI, shifting the question from whether to use AI to how to implement it effectively.
The Harness: More Than Just the Model
A large language model alone is insufficient; it must be surrounded by a harness—the framework, tools, memory systems, safeguards, and knowledge grounding that transforms raw capability into a practical product. LangChain provides this harness layer that makes models genuinely useful.
Agentic Systems: The Real Breakthrough
Modern agents combine grounded information, tool use, managed memory, safeguards, and iterative problem-solving. Models like Claude 3.5 Sonnet and OpenAI o1 demonstrated that when capability reaches a threshold, agentic systems become viable—representing the true inflection point, not just model improvements alone.
Why Open Systems Matter for Enterprise AI
AI as Fundamental Technology Requires Specialization
AI can only be useful across diverse domains if companies can build specialized, domain-specific systems. Scientists, designers, roboticists, and enterprises each need AI tailored to their unique problems—not one-size-fits-all solutions. This requires open tools and frameworks.
The Flywheel Effect: Continuous Improvement
Specialized AI systems improve over time through use—the more a company uses its agent, the smarter it becomes, driving more usage and further improvement. This flywheel only works when companies own and control their systems; outsourcing specialized intelligence to third parties breaks the cycle.
Intellectual Property Is Intelligence
Every company's competitive advantage is built on specialized, domain-specific intelligence. Outsourcing this intelligence to external providers is fundamentally incompatible with maintaining competitive advantage. Companies must build and control their own specialized AI systems internally.
Future Companies Built on Harnesses, Not Processes
Today companies are built on business processes; tomorrow they will be built on AI harnesses. LangChain will become the operating system that enables every company to create specialized, autonomous workflows that continuously improve and adapt.
Model Performance vs. Cost: The Open Weight Advantage
Nemotron 3 Ultra Reaches Frontier Performance at Fraction of Cost
Open weight models have crossed the capability threshold. Nemotron 3 Ultra achieves 86% on internal benchmarks (compared to Claude Opus at 87%) while costing 10 times less, demonstrating that open models now deliver frontier-level performance at dramatically lower cost.
Cost Efficiency Enables Broader Exploration
When intelligence is cheap and fast, companies use more of it and explore larger solution spaces. Nemotron's computational efficiency allows agents to iterate quickly, test more possibilities, and find better answers—similar to how faster human thinking enables better problem-solving.
Balancing Frontier and Open Models
Start with frontier models to understand potential and validate approaches quickly, then specialize with open models for domain-specific tasks. Both have roles: frontier models for general capabilities and continuous improvement, open models for specialized, cost-effective sub-agents.
Building Specialized Super Agents
Specialization Requires Three Layers
A capable base model is necessary but insufficient. Specialization requires combining the model with a customized harness (tailored prompts and tools for that domain) and access to proprietary domain-specific knowledge and information. Each layer can be independently optimized.
Real-World Example: Supply Chain Optimization
NVIDIA builds specialized sub-agents for hard optimization problems like supply chain and chip design. These super-agents use Nemotron 3 Ultra with LangChain, connected to proprietary tools and knowledge. They're not general-purpose; they're built for one specific job with dedicated refinement teams.
When to Specialize: The Trigger
Specialize as soon as frontier models are good enough for your use case. For many applications, frontier models improve continuously and never need replacement. Specialize only when you need domain-specific capabilities, proprietary knowledge integration, or cost optimization for repeated tasks.
Post-Training: The New Frontier
Once a harness is built and working, companies can now post-train the model specifically against that harness to improve performance further. This capability—continuously improving the model for a specific workflow—has never existed before and represents a complete breakthrough in enterprise AI.
The Complete Stack: From Model to Runtime
Building Blocks of the Agentic Stack
Creating agents requires: a world-class language model, a framework (LangChain Deep Agents), domain-specific knowledge and tools, memory systems, guardrailing, fine-tuning capabilities, and a secure runtime. Each component must be integrated and work together seamlessly.
Security and Access Control: Non-Negotiable
Deploying agents requires solving security and access control first. Just as employees need role-based access to files, networks, and tools, agents need granular access control. IT organizations must be able to onboard, sandbox, and govern agents like they do employees.
The Blueprint Approach: Reducing Complexity
Tools are still arcane and complex. Blueprints package all key ingredients together—model, framework, knowledge integration, memory, guardrails, fine-tuning, and runtime—into a single, deployable template. This dramatically reduces friction for enterprises building their first agents.
Deploy Anywhere: Cloud, On-Prem, or Local
The complete stack now runs everywhere: in the cloud, on-premises, on DGX stations, or even on local hardware. Companies can choose their deployment environment based on security, latency, and cost requirements without changing the underlying system.
Agents as Tools, Not Humans
Agents Are Software, Not Conscious Entities
Agents are tools like dishwashers or autonomous lawnmowers—software that automates tasks humans previously did. They have no consciousness, no awareness, and no biological properties. Understanding how they work is essential to improving them, which is why transparency and explainability matter.
AI Adoption Creates New Jobs, Not Fewer
As companies deploy more AI, they hire more people—not fewer. Software engineers shift from writing code to building agents, creating evals, benchmarks, and guardrails. The work required to integrate AI into organizations is substantial and creates entirely new roles and skill sets.
Evals and Subject Matter Experts Drive Improvement
Quantifying agent performance requires domain experts who understand the business context. Subject matter experts inside enterprises provide feedback, create evals, and work with systems to automate tedious work while focusing on creative, intellectually stimulating tasks.
The Future: From Automation to Amplification
Current Approach: Automating What We Did Before
Most current agent use cases focus on automating existing processes—what companies already do. This is valuable but limited. The real unlock comes from asking what we couldn't do before that we now can do with AI assistance.
Future Approach: Enabling New Possibilities
The true potential of agents lies in amplifying human capability to do things previously impossible. This requires ambition and imagination—thinking beyond automation of existing workflows to entirely new capabilities and workflows that AI enables.
Today's Announcement: All Pieces Now Available
NVIDIA is providing all foundational building blocks needed to build domain-specific super agents: Nemotron 3 Ultra model, LangChain Deep Agents framework, blueprints for easy deployment, and OpenShell secure runtime. Every developer and company now has no excuse not to build specialized agents.
Notable quotes
The last six months changed everything. Now finally AI is useful. — Jensen Huang
Every company is built on intelligence. You can't possibly not continue to control it. — Jensen Huang
In the future, most companies will be built on harnesses, not business processes. — Jensen Huang