LLMorAgent
A language model is a part: you call it, it produces text, and there it stops. An agent is the system that calls that part in a loop, carries out what it proposes and calls it again with the result, until the goal is reached.
Point by point
| Criterion | Floor 1Floor 1 · The Modela solid block on its own: the prediction machineLLM | Floor 3Floor 3 · The Agentthe block caught in a loop: it starts again until it gets thereAgent |
|---|---|---|
| What is being named | A part: a model you call | A system that calls that part in a loop |
| What comes out | Text, nothing else | Actions, and their effects |
| What stops it | The end of its answer | The goal reached, or a limit you set |
| Calls to the model | One | One per turn, as many as it takes |
| Cost of a unit of work | Known: the tokens in and out | Unknown in advance: it depends on the number of turns |
| What can go wrong | A false sentence | An act performed in the world, sometimes irreversible |
On the ground four situations
In a processing chain, one step sends a paragraph to the model and gets a category back.
Floor 1 · The Modela solid block on its own: the prediction machineLLMOne call, one output, done: that is the whole extent of what a model does. What comes before and after is code someone wrote.
Taking over a codebase: read the files, attempt a change, run the tests, read the error, start again until everything passes.
Floor 3 · The Agentthe block caught in a loop: it starts again until it gets thereAgentThe model is called at every turn, but what drives the task is the loop around it: the loop is what calls again, with the result of the previous turn.
A model returns an action plan in ten numbered steps, precise and convincing.
Floor 1 · The Modela solid block on its own: the prediction machineLLMThe useful counter-example: describing steps is not carrying them out. As long as nobody reads that plan in order to run it, there is only text.
A conversation interface that answers, keeps your preferences and can open a document you point it at.
AssistantNeither one: the model alone keeps nothing and opens nothing, and nothing here loops towards a goal. Everything described belongs to the harness, and the product is called an assistant.
The test that settles it
Take a task and count the calls to the model in the execution trace. A single call: you were looking at a model, and all the rest is code written around it. A series of calls, each starting from the result of the previous one: that is an agent.
The trap
“The AI decided to refuse the application.” The sentence points at the model and blames the wrong party: a model decides nothing on its own, it produces text when called, then it stops. What decides is the system that chose to execute that text rather than put it to someone.
Check click your answer
Level 1 · Recognise
You call a language model, and you do nothing more. What happens next?
Level 2 · Distinguish
An agent works for half an hour on a task. How many times was the model called?
Level 2 · Distinguish
Two products call exactly the same model: one returns an answer and stops, the other carries a task through by chaining dozens of steps. Where does the gap come from?