An LLM is an engine, not a car.
A large language model is an engine. On its own it is stateless, goalless, and blind to your systems. The harness built around it is the car, and that is where the engineering lives. A person is still the driver, and still sets the destination. Confuse the engine for the car and you will keep being surprised by what AI can and cannot do.
The model turns an input into a plausible output. That is the whole job.
A language model does one thing: it takes text in and produces the next text out, one plausible piece at a time. It is very good at that. But three things are true about it that people forget the moment a demo impresses them.
It is stateless. Each request starts cold. The model does not remember your last conversation, your account, or what it did an hour ago unless something outside the model feeds that back in. It is goalless. It has no plan for your business and no stake in the outcome. It answers the prompt in front of it. And it is blind. It cannot see your CRM, your inbox, or your files. It only knows what is placed in front of it in the moment.
None of that is a flaw. It is what an engine is. An engine does not know where you are going. It converts fuel into motion when you ask. Everything that makes the motion useful is built around it.
Three parts. Only one of them is the model.
It helps to name the pieces plainly. When people say "the AI did that," they almost always mean the car did, not the engine.
| Part | What it is |
|---|---|
| The engine the model | Turns an input into a plausible next output. It has no memory of yesterday, no goal of its own, and no way to see your systems. Powerful, and inert on its own. |
| The car the harness | Everything built around the model so it can do work: the prompt it gets, the tools it can call, the memory it carries, the limits on what it may touch. This is where the engineering lives. |
| The driver a person | Sets the destination and stays accountable for arriving. The model does not decide where the business is going. Someone still does. |
The harness is where the engineering lives.
Give ten teams the same model and you will get ten different results. The difference is not the engine. It is the car built around it: how the prompt is assembled, which tools the model can reach, what it remembers between steps, and the limits on what it is allowed to do. That work is invisible in a good demo, which is exactly why the demo looks like magic.
This is also why the model itself is becoming a commodity. The engines are strong, they keep improving, and they are converging on the same shape. The advantage does not sit in the engine you pick. It sits in the car you build around it and who is allowed to drive.
When people say the AI did something impressive, they almost always mean the harness did. The engine just ran.
Wrap the engine in a loop and you get an agent.
An agent is what you get when the harness runs the engine in a loop: read the current state, decide the next step, take an action, look at the result, and go again until the goal is met. The model supplies each next step. The loop, the tools, and the limits around it are what turn a string of plausible outputs into work that actually moves.
That framing matters because it tells you where to put your attention. You do not make AI trustworthy by waiting for a smarter engine. You make it trustworthy by building a better car: one where the goal is declared before work starts, every action is recorded, and the high-impact moves route to a person. The engine got you motion. The rest is what keeps you on the road.
Common questions
What is the difference between an LLM and an AI agent?
An LLM is the model: a component that predicts the next output from an input. An agent is the whole system built around it, a model wrapped in a loop that can read state, call tools, and act toward a goal. The LLM is the engine. The agent is the car with a driver in it.
What is a harness?
The harness is everything around the model that turns it into something useful: how the prompt is constructed, which tools it can call, what it remembers, and the guardrails on what it may do. Most of what people credit to a smart model is actually the harness doing its job.
If the model is a commodity, where does the advantage come from?
From the harness and the governance around it. Two teams pointed at the same model get very different results depending on how the loop is built, what it is allowed to touch, and whether a person can see and steer it. That is the part you own, and the part worth investing in.
Does a better model remove the need for a harness?
No. A stronger engine still needs a car and a driver. Better models raise the ceiling on what the harness can do, but they do not connect to your systems, remember your context, or decide your goals on their own. The work around the model is what makes it safe to run.
Keep reading
If the engine is a commodity and the car is where the value sits, the next question is who builds and owns that car. That is the subject of the next explainer.
Explainer · The Stack Bring your own agent Why the agent you run is becoming a commodity, and where the real value sits: the governance layer above the tool. Read it →Want the car built around your engine?
Book thirty minutes and we will map where a governance layer and a deployed crew would take the most work off your plate, on the agent you already run.