The Model Is an Ingredient. What Should a Company Own?

Ingredient and harness

The model is not your company’s advantage.

That sounds wrong at first. Frontier models are expensive to build, difficult to reproduce and increasingly capable. They can reason, write, code, search and work across tools. A company without access to that intelligence may be operating at an obvious disadvantage.

But access to the same model is available to everyone else.

In a conversation published by LangChain, Jensen Huang described the large language model as “the essential ingredient” and “the essential enabling technology.” Then he made the distinction that matters: to turn it into a useful product, “you have to surround it with what is now known as a harness.”[3]

An ingredient can be essential without being the finished system.

Huang’s description of an agentic system moves beyond the model. He names grounding in information and knowledge, tool use, managed memory, safeguards and the ability to iterate until the job is done. Later, he adds the knowledge graph, fine-tuning, the harness itself and a secure runtime with access control.[3]

AGL’s reading begins with a practical consequence: if the model is one component, company-specific value must be built somewhere else too.

Company capability

A frontier model gives the company general intelligence. The harness gives that intelligence a way to act. Company knowledge, tools and workflows give it a particular job. Permissions and safeguards define what it may touch. Memory and evaluation allow the system to improve through use. Human authority determines where action stops and responsibility returns.

This is how a capable model becomes a company capability.

The difference matters because companies often speak about AI adoption as though choosing a provider were the central decision. It is important, but it is not enough. Two companies can buy access to the same model and build radically different capability around it.

One gives the model occasional prompts. The other connects it to a defined workflow, supplies the right operating knowledge, gives it bounded access to tools, records what happens, tests representative cases and carries corrections into the next cycle.

The underlying intelligence may be shared. The operating capability is not.

Huang makes this point through specialisation. In the LangChain conversation, he describes company workflows as proprietary and important, and suggests that specialised super agents built around them could become a company’s “crown jewels.” He also says frontier models will continue to be used “absolutely, and tons of it.”[3]

This is not a choice between frontier intelligence and proprietary systems. It is a question of where each belongs.

Use frontier models for the general capability they provide. Build specialisation around the work only your company does in its particular way.

That specialisation may include customer and operational knowledge, but it is not merely a database. It also lives in decisions and exceptions: how a lead is qualified, when an unusual order needs human attention, what evidence is sufficient for a claim, which supplier risk changes a recommendation, what can be published, who may commit spend, and which result counts as complete.

A workflow becomes genuinely company-specific when those judgments are made explicit enough for a system to act on them without pretending that responsibility has transferred to software.

This is AGL’s extension. The valuable asset is the operating layer around the model: context, knowledge, workflows, memory, tools, permissions, correction history, evaluations, safeguards and documented human authority.

What belongs to you

Huang’s line near the end of the conversation gives the ownership claim its shortest form. Speaking about domain-specific super agents, he says: “They belong to you.” He continues: “You build them, you improve them, you refine them over time.”[3]

“Belong” should not be treated as a decorative word.

If a company builds and improves a specialised agent through its own knowledge and use, ownership must mean more than access to a vendor account. The company should be able to identify the assets it controls, understand where they live, recover them, govern who can change them and continue the capability when a component or provider changes.

That does not mean the company owns the third-party model underneath the system. It does not mean every component must be open-source. It does not mean a model can be exchanged instantly without engineering cost.

It means the company-specific layer should not vanish merely because the intelligence provider changes.

The test is simple to ask and difficult to pass:

If your current provider disappeared tomorrow, what part of the agent would survive?

Would the company retain its knowledge base, workflow definitions, prompts and policies, tool configuration, permissions, evaluation cases, logs, memory, correction history and documentation? Could a different model be connected without rebuilding the company’s operating world from the beginning? Would staff know how to restore the system and where human approval remains mandatory?

A company that cannot answer those questions does not yet own a durable capability. It may have an excellent AI service. It may even have useful automations. But the specialisation is still inseparable from the provider through which it was assembled.

The aim is not to move everything in-house. Huang’s own account is complementary: general models remain available through the cloud, while companies build specialised capabilities on top.[3] The sound architecture uses external intelligence without outsourcing the company’s identity, knowledge and operating development along with it.

That is why Huang’s other statement matters: “Your company’s intelligence is who you are.” His conclusion is that a company must continue to control, improve and strengthen that specialised intelligence rather than simply hand it away.[3]

AGL agrees with the ownership implication while making its own boundary explicit. An agentic system can carry knowledge forward and act within approved boundaries. It does not inherit the company’s moral or legal authority. Consequential decisions remain with accountable people, and wider agent authority should be earned through visible, verified performance.

The model will change. The provider may change. The interface certainly will.

What your company learns through building, correcting and operating the system should remain part of the company.

Use the frontier. Own the specialisation.

Continue through the Company route →

Source notes

1. Jensen Huang, “Jensen Huang: Why companies need open agent systems”, published by LangChain on YouTube, 8 July 2026. Quotations were checked against the available transcript; the source is a direct recorded conversation and also contains product-specific advocacy for NVIDIA and LangChain systems.

Independent AGL editorial commentary. No affiliation, endorsement, partnership, sponsorship or technical parity is implied.