An agent: model, tools and a loop

The word "agent" is often used for anything that involves AI. In this book we use it in a narrow sense that can be checked. An AI agent is a large language model given tools, running in a loop, to pursue one goal until its stop condition holds. There are five parts, and you can point to each of them in any agent:

1. Model. A large language model that reads text and writes text. It has no hands; it can only write requests.
2. Tools. Functions the app provides: read files, search, run tests, create drafts. The model writes "call this tool with this input", then the app carries it out.
3. Loop. The tool result goes back to the model, the model picks the next step, and the cycle repeats. This is what sets an agent apart from a single answer.
4. Goal. A sentence that says when the work counts as done, for example "every test in this folder passes" or "a two-page summary with file references".
5. Stop condition. Done, a turn limit, a cost limit, or a point where the agent must ask a person.

If one part is unclear, the trouble almost always shows up there. A vague goal makes an agent stop too early or never stop. A tool that is too powerful turns a small mistake into big damage. A loop without limits burns money. Tool use itself is covered in the lesson on tools and connectors; here we look at what happens when that call is repeated many times by the model itself.

Three things are often confused: a chatbot, a fixed workflow and an agent.

• A chatbot answers one message with one answer. You decide the next step. It may use one tool, such as search, but it does not plan a chain of steps itself.
• A fixed workflow is a chain of steps whose order a person wrote: fetch the e-mail, summarise, save to a folder. A model may be used in one step, but the path never changes. Easy to test and easy to audit.
• An agent picks its next step itself based on the result of the previous one. Its path can differ every run, which is why it can handle tasks that cannot be ordered in advance, and also why it is harder to oversee.

The difference is not about being advanced or not. A fixed workflow is often the better choice, because it is cheaper and easier to check. An agent earns its place when the steps truly depend on what is found along the way, for example fixing a failing test: you do not know which file to open until the error message has been read.

To make it concrete, look at the loop of an agent asked to fix a failing test. Turn one: the model reads the goal and calls the "run tests" tool. The app runs it and returns the output: one test fails because a date is written in the wrong format. Turn two: the model calls the "read file" tool for the function that formats dates. Turn three: the model calls the "edit file" tool with a fix. Turn four: the model runs the tests again, and this time they all pass. The stop condition "all tests pass" holds, and the agent writes a report.

Notice two things. First, every turn adds to the context: the goal, every tool request and every result are sent to the model again. That is why long tasks get more expensive with each turn. Second, the report at the end is the model's writing, not evidence. The evidence is the test output in turn four. In chapter three we learn to read traces like this.

A small exercise for this week: take one task you often do with AI and write its five parts on one sheet. If the "loop" part is empty because you decide every step yourself, that is fine: it means what you need is a good prompt or a fixed workflow, and you save cost and risk. If the steps really change depending on what is found, write the goal and stop condition first, before granting any tool. The next lesson covers how to choose among the three.