What is an AI agent?

The short version

You ask AI to prepare a file and receive instructions to follow yourself. You ask again and receive a file you can download. What changed? An app that answers your question does not necessarily have the tools to carry out the work you describe.

In the episode, we upload an order file to Claude and request an Excel file listing unpaid amounts and reminder messages. With file creation disabled, we receive a table in the chat and instructions to copy it. With the capability enabled, in a fresh chat using the same file and request, Claude reads the file, calculates the amounts and creates a new file. This introduces an agent: a model directing tools through a task, while you remain responsible for checking the result.

Key terms

• Model: the system that processes your request and context and generates a response; it can also select a tool when the app makes one available.
• Tools: capabilities through which an app performs actions, such as reading a file, calculating or searching the web.
• Agent: a system in which a model uses tools and chooses subsequent steps based on their results to pursue a goal.
• Effort: a setting, where available, that affects how much effort the model applies to processing a request.
• Tokens: the units a model uses to process text; reasoning may consume tokens in addition to the input and visible reply.
• Permissions: the access and actions you allow the app, such as editing files or sending messages.

The explanation

The same clerk, now with tools

Continue the series' government-office analogy: you submit a request at a counter, and the clerk behind it is the model. Without tools for the job, the clerk can explain what you should do, but leaves you to do it.

Give the clerk a ledger and a calculator, and they can open the ledger and calculate. Those are tools. Ask for the whole job, and let the clerk choose actions and use their results to decide what comes next: that illustrates an agent.

The episode's shorthand, “Model + Tools = Agent,” is a memory aid. Having a tool alone does not make every interaction an agent; the point here is that the model directs execution as the task unfolds. Anthropic's guide to agents distinguishes predefined workflows from processes directed by the model.

The experiment: a table to copy or a file to open?

The kitchen file contains 12 orders, with fictional names and numbers: seven paid and five unpaid. The requested result is an Excel file with unpaid customers, the amount each owes and a short reminder message.

On the first attempt, file creation and editing were disabled. Claude displayed a table in the conversation and asked the user to copy it into Excel. On the second, the capability was enabled, and the exact same file and request were sent in a fresh chat. The recording shows file inspection, calculation and file creation, followed by a result that can be downloaded and opened.

That is what happened in the recording, rather than a guarantee for every attempt. Setting names, locations and availability can change; consult Claude's file-creation guide if your interface differs. Distinguish text describing work from tools actually producing the requested file.

A polished file can use the wrong column

The recorded result contains these unpaid customers, with their names translated here:

Check Nadia's source row: 5 boxes × 120 = 600. Then add the five amounts: 360 + 300 + 500 + 600 + 250 = 2,010.

But Nadia's order number is 9. If the system uses that instead of the quantity, it calculates 9 × 120 = 1,080. The arithmetic is correct, but the chosen column is wrong. Repeating that mistake for all five orders produces 6,890 instead of 2,010. The episode illustrates this possible failure; Claude did not make it in the recorded result.

Coding, reviewing and working as a team

The episode names Claude Code, Codex and Google Antigravity as examples of coding tools using the agent idea: reading and editing project files within their granted access. Instead of copying code out of a chat, you can receive changes in the project itself.

Work can also be divided: one agent writes and another reviews. Or one summarizes emails and another drafts replies, with you reviewing before sending. These are examples of organizing work, rather than guarantees that a reviewer catches every error or that every app automatically sets up a team.

Higher effort does not guarantee a better result

Where an Effort setting is available, deeper processing can consume more tokens and take longer; Claude's effort-settings guide explains that trade-off. Match the setting to the task and compare correctness and time, instead of always choosing the highest level.

The episode's thirsty person, stuck deciding which cup to use, illustrates the point: additional deliberation helps only when it brings you closer to the goal. The animated token numbers and dial are illustrations, rather than measurements of the experiment.

Where do tools run, and what can they do?

In the file experiment, execution happens in a computing environment provided by Claude, rather than inside Excel on your device. Downloading the finished file does not mean the data stayed on your device throughout the task.

Before uploading, ask whether you own the file or have permission to share it, whether it contains real names, phone numbers or sensitive information, and what work you want to authorize. That is why the experiment uses fictional data.

An agent may also work with files on your own device. That does not automatically make all processing local: the app may send content to a model service. Check how the tool works and grant only the access the task needs. Drafting reminders does not require permission to send them, and preparing a shopping comparison does not require buying anything.

Try it yourself

Order number,Name,Item,Quantity,Unit price,Payment
1,Mona,Chocolate cake,1,250,Paid
2,Karim,Basbousa,2,180,Unpaid
3,Salma,Kahk,3,120,Paid
4,Youssef,Cheesecake,1,300,Unpaid
5,Hala,Kahk,2,120,Paid
6,Omar,Chocolate cake,2,250,Unpaid
7,Reem,Basbousa,1,180,Paid
8,Tarek,Cheesecake,2,300,Paid
9,Nadia,Kahk,5,120,Unpaid
10,Sami,Basbousa,1,180,Paid
11,Laila,Chocolate cake,1,250,Unpaid
12,Fadi,Cheesecake,1,300,Paid

Create an Excel file from the attachment listing the customers who have not paid yet, the amount each owes, and a short, polite reminder message for each customer.

I want to buy Bluetooth headphones, and my budget is X. Search three online stores for the best options within that budget. Compare prices, specifications and delivery costs, and make a table of the best three headphones with a purchase link for each. Finally, save the comparison in an Excel file. Do not buy anything.

Open the links and check prices, specifications and delivery costs yourself at the time of the exercise. If a store is inaccessible or a delivery cost is unavailable, the result should identify the gap instead of inventing a value. A comparison does not authorize a purchase.

Why this matters for building software

When you build software by talking to AI, define the result you expect to receive: working changes in the files, rather than just an explanation of how to make them. State the task boundaries and prohibited actions, then follow execution and try the result.

Instead of “add a button,” you could ask: “Add a button that saves the order, show a success message after saving, and test that the order appears in the list.” The agent can choose implementation steps, but you still need evidence that the button saves and the data appears where expected. More tools or reviewers do not replace running and testing the software.

Related lessons

• What is AI?: the difference between a model and a chat app.
• How do you ask for exactly what you want?: specifying the task, constraints and expected result.
• Why does AI forget what you told it?: the working context the model receives.
• Why does AI lie so confidently?: why evidence and result checks matter.
• Why so many AI models?: choosing for the task, speed and cost.