What is the difference between an AI model, a tool and an agent?

Engine and car

The engine delivers the raw capability: processing language, recognising patterns, predicting what comes next. Everything else is delivered by the car.

The car uses the engine, but it also has a body, a steering wheel, seats.

A tool takes the same engine and builds a purpose around it: a text generator, an image analysis, a summary button in the accounting software that has been in the building for years. Developers now often call this car the harness, the environment that hands the model files, a terminal and tools. The name is newer, the layer is the same.

The four layers

Anyone buying should know which layer they are negotiating on. Four layers are enough, each with a case that almost every company knows.

You can picture an API as a kind of standardised power socket, or as a menu with a waiter in a restaurant.

Between the interface and the agent sits a middle step that many companies already use without calling it that. Automation platforms connect services through exactly these interfaces, by dragging and dropping instead of writing code. That is not yet an agent, because the order of steps stays hard-wired. It is a car with cruise control.

• Model: the capability itself. In practice: two proposals written in the same tool but by two different models. The gap in quality is immediate, although nobody changed the surface.
• Tool: the purpose around it. In practice: the button in the customer portal that summarises a request. Nobody calls it AI, everybody uses it.
• Interface: ordering without seeing the kitchen. In practice: the inventory system sends a product description out through an api and gets a corrected version back, without anyone opening a chat window.
• Agent: the fleet under management. In practice: an incoming complaint is read, sorted, enriched with your own knowledge and handed over as a draft reply.

The fleet and its project manager

One car is not enough for an agent. It needs a delivery van for transport, for example a tool that turns a photo into text. It needs a fast car for the creative sprint, which is the language model itself. And it needs a navigation system, which is search or the internal knowledge base.

What makes a super agent special is that, like a kind of project manager, it coordinates these different tools and models and data sources.

The example in the course is an incoming email. The agent reads it, sorts the request, pulls the answer from the internal knowledge base in a standard case and writes a draft. If the case gets harder, it gathers what is known, proposes a solution and hands over to a person. That is exactly the boundary I still draw today.

Two things separate an agent from a fixed workflow. It decides the next step itself instead of running a hard-wired chain. And if a database sits behind it, it remembers what it found out and uses that again next time. Providers agreed on rules for this. The Model Context Protocol works like an interface, but it also states what a tool actually does, so a model can handle it more easily. A second protocol governs how two agents talk to each other.

Why the layers matter when you buy

The tool is what your people see and operate. For risk and law, though, the properties of the model underneath decide: what it was trained on, what it can do, where it reliably gets things wrong. So it pays to look under the bonnet before a signature goes underneath. Anyone who only knows the name of the surface cannot answer that question.

When I recorded the module, that protocol was a side topic for tinkerers. By now it is a common way for an agent to reach its tools. The layers have not moved because of it. What got better is only the connection between them.

What follows for an AI department

An AI department is this order applied to one department. Every agent gets a role, tools and access, and a person approves. Knowing the layers gives you the better questions in a sales conversation: which model sits underneath, what purpose does the tool serve, which tools may the agent operate, who sees the draft before it is sent.

How I split the roles across several models is in Who thinks, who works, who just runs. What such a department costs and what the price is made of is in What does an AI department really cost?. How far your own assistant may already go is settled by the Jarvis ladder. The way from the first department to daily operation is described on the page about AI implementation.

Sources and links

This note keeps growing

2026-09-03: Planted from module 1.5 of a certification course on the EU AI Act, 2025.