What is a foundation model?

A foundation model is a large AI model trained on broad, general data so that it can be adapted to many different tasks, from drafting text to reading documents to writing code. Most AI agents are built on one. The foundation model supplies the general ability to read and reason; the rest of the system decides what that ability is used for and what it may touch.

Why does the foundation model matter once agents act?

Its strengths and weaknesses pass straight into every agent built on it. If the model tends to produce confident errors on certain kinds of question, so will the agent. If it can be swayed by instructions hidden in a document, so can the agent. And because the same model may sit under dozens of agents, a change to it touches all of them at once.

That last point matters in regulated work. Model providers update their models. An agent certified to handle complaints on one version may behave differently on the next, even though nobody at your firm changed anything.

How are foundation models used in practice?

Few firms train their own. They call a model provider's model, or host one, and build agents around it with instructions, sources, tools and memory. Some fine-tune a model on their own data. Many route different tasks to different models, which is the job of a model router.

In each case the model is a component. The controls that matter in a regulated business, such as which customers an agent may contact or what amount it may refund, belong outside the model. That is why we argue model independence is a governance requirement: if the rules live inside one model's prompt or weights, changing model means rebuilding your controls.

What does that look like in an example?

A consumer lender runs three agents on the same foundation model: a complaints triage agent, a collections drafting agent and an income verification assistant. The provider releases a new version. The lender re-runs certification for each agent's capabilities against its own cases before switching. The triage agent passes. The collections agent now drafts letters with a different tone that fails the adversarial gate on two cases involving customers in difficulty, so it stays on the old version until its brief is adjusted.

How is a foundation model different from a reasoning model?

A reasoning model is a foundation model trained or set up to work through intermediate steps before answering, often at higher cost and slower speed. Every reasoning model is a foundation model; not every foundation model is built to reason at length. For agents, the choice is task by task, which is another reason to keep the rules outside the model.

Why can't the foundation model govern itself?

Because the same system cannot be both the actor and the check on the actor. A model asked to judge its own proposed action brings the same blind spots to the check that it brought to the action. The argument is set out in why the model cannot be its own supervisor.

Where Pryme fits

Pryme sits between foundation models and the business, and works with any of them. The model proposes; the platform decides what each agent may do, based on its certificate and bounds. Changing the model does not change the rules. See the governed agentic operating system, explained.