Lesson 1 of 5
What has to change for AI to carry out work?
In module 1 a chat application produced the wording of a garden reply and nothing else happened: nobody sent the message and nothing checked the weather. Module 2 showed what software can do beyond wording, such as read files, run programs and send requests, and that it does so only within the access it was granted. This lesson joins the two. It explains what has to be added around a model before a request such as “compare these documents and save the result” produces a saved file, and it names the parts so the rest of the module can trace them.
Numbered links identify sources. Underlined terms offer a definition without leaving the lesson.
Generating text does not save a file
Start with the request this module follows. The garden organizer keeps documents in a folder named Garden. Inside it, a folder named Suppliers holds quotes for a new fence along the north boundary. The organizer types into an that passes typed text to a and shows the reply, and nothing else: “Compare the fencing quotes in Garden/Suppliers and save a comparison table as fencing-comparison.txt.” Everything about this example is authored for the lesson.
Module 1 explained what the model does with that request: it generates a reply one token at a time from its learned values and the it was given. The context here contains the request and nothing from the Suppliers folder, because this application supplies nothing else. The reply therefore has no evidence from those quotes. Any figures it supplies would be unverified guesses, even if a guess happened to match. The reply might still be fluent. It might even end with the sentence “Saved as fencing-comparison.txt.” Module 2 explained why that sentence cannot be true here: a changes only when a running performs a write operation on it, and this application performs no such operation. The sentence is a generated claim of the kind module 1 taught you to distinguish from a receipt.
Two gaps separate this request from its result. Information has to come in: something must read the quotes and put their contents where the model can use them. An effect has to go out: something must write the table to storage. Both are operations, and module 2 established that operations are performed by software running under some granted access. A model can describe an operation in words. It cannot perform one. OpenAI's documentation for developers building on its models states the division plainly for the tools a developer defines: the model produces a request naming an operation and its inputs, the developer's own code runs the operation, and that code passes the result back to the model. 1
A tool is an operation the surrounding software offers the model
In this course a is an operation that the software around a model makes available for the model to request. A tool has a name, the inputs it expects and the kind of result it returns. Listing a folder, reading a file, writing a file, fetching a web page, sending an email and running a program are all tools in this sense. Each is an ordinary operation from module 2: a file operation, a process started on some , or a request sent by a to a server. Many tools are the of module 2 wrapped so that a model can ask for them. The word is used more loosely elsewhere; this site's tools directory, for example, lists whole products. Here it means one operation.
The model asks for a tool by producing a : text in an agreed format that names the tool and gives its inputs, which documentation often calls arguments. Here is an authored illustration of the idea. No real product uses exactly this wording, and the formats differ between products:
call: read_file
input: Garden/Suppliers/northfence-quote.txt
The software recognizes that text as a request rather than as part of a reply, runs the read, and gives the result back as more context:
result: Northfence. 40 m timber, £2,150 including installation. Delivery 3 weeks. 5-year guarantee.
Nothing new has happened inside the model. It generated tokens, as module 1 described, and the surrounding software treated a particular pattern of tokens as a request. The model's interface supports output in the tool-call format. This still does not execute the operation. Until software runs it, a tool call is a proposal. OpenAI's documentation describes the same sequence for the tools a developer defines, and it separates them from tools the provider hosts itself, such as web search. The division is the same in both cases: the model requests, software executes, and the result returns. 1
A harness organizes model calls, tools and execution
The software that does all of this has a name in this course. A is the software that sends the model its context, recognizes tool calls in the model's output, decides whether to run them, runs the ones it allows, returns each result into the context, and repeats until the work stops. In module 1's vocabulary a harness is an application, or the part of one, with this particular job. Not every application is a harness. A chat window that only shows replies is not one, and the plain application at the start of this lesson is not one either.
The harness is also where the access lives. It holds the connections and from module 2, it decides which tools exist at all, and it keeps the record of what ran. The site keeps a directory of harnesses. Its entry for Claude Code, Anthropic's command-line coding product, described it on 29 September 2026 as the harness developers use to build and run coding agents from the terminal, and listed capabilities such as reading a whole codebase and carrying out version-control work through typed requests (tool entry). Read such a list with this lesson's word: each capability is a tool, or a set of tools, that the harness exposes.
One more piece of module 2 explains where things happen. A harness may run as a process on your own laptop while the model runs on the computers of an access provider, the role module 1 described. The harness is then a client sending requests to the model's API and reading the replies. When someone says “the agent read my files,” the precise statement is that a process on their host, running the harness program, performed a read because the model asked for it. The available tools and their granted access set the boundary on operations. The model's choices determine which allowed operations it requests.
An agent is the whole arrangement, not a different kind of model
Module 1 closed by promising an explanation of agents. An , in this course, is a system in which a model repeatedly chooses operations, a harness carries them out, and the results come back to the model to inform its next choice, until the work is judged finished or is stopped. The agent is the arrangement: model, harness and tools, together with whatever rules govern them. A common misconception is that an agent must be a different kind of model. The same model can write a reply in a chat window and, placed in a harness, propose tool calls; what changed is what surrounds it. Module 1's comparison of three applications using one model is the same pattern. Models do differ in how reliably they handle tool use, and choosing between them is a later module's topic. The difference this lesson needs is the one between a model and an arrangement.
The Hugging Face Agents Course, a free online course from a company that hosts models and datasets, introduces the same arrangement as a cycle of thought, action and observation. Its example asks an agent about the weather: the model decides that it needs current data, produces a call to a named weather tool with a location as its input, the tool's answer comes back, and the model then writes a reply that uses it. 2 That example is a teaching illustration, as ours is. The next lesson traces the cycle in full and explains which parts of such a record are evidence.
Carry this forward
- Generating text describes an operation; it does not perform one. A saved file needs a process to write it, and a generated sentence saying “saved” is a claim.
- A tool is an operation the surrounding software offers the model. A tool call is the model's generated request to use one, naming the tool and its inputs, and it remains a proposal until software runs it.
- A harness supplies context, recognizes tool calls, runs or refuses them, returns results and keeps the record. The harness, not the model, holds the access.
- An agent is the arrangement of model, harness and tools working in a loop. The same model can run in a chat window or in a harness.
Lesson check
Choose an answer for each question.
This is a practice check. Answers reset when you leave this page. Results are not saved and do not place you in another module or award a qualification.
The parts have names. The next lesson follows one request through them, step by step, and shows which lines in the resulting record are evidence of what happened.
Sources · Reviewed 2026-09-29
- 1. OpenAI: Function calling
- 2. Hugging Face Agents Course: Understanding AI agents through the thought-action-observation cycle
Sources support nearby explanations. Product details can change. Scenarios and model results are authored illustrations unless a lesson explicitly identifies a recorded execution. Module 7 identifies its offline code checks.