Lesson 3 of 6 · Agents

Claude Code and Codex

Two names you keep hearing. Both are agents built for programmers, both are included in subscriptions you might already pay for, and both can do useful work for people who have never written a line of code.

AI Basics lesson 3 listed Claude Code and Codex under "coding harnesses" and moved on. This lesson stops and looks. They are the most widely used agents you can sign up for today, they are the clearest example of the four ingredients from lesson 1, and if you have a paid ChatGPT or Claude plan, you already own one. Everything here is as of August 2026, verified against the makers' own pages on the day this was written. Expect names and prices to drift.

What they are

Claude Code is Anthropic's agent. Their own one-line description: it reads your codebase, edits files, runs commands, and integrates with your development tools. A "codebase" is just a folder of files that happen to be a program. Strip the programmer words and it says: it reads a folder, changes files in it, runs things, and talks to your other software.

Codex is OpenAI's agent. Their description of the command-line version: inspect code, make changes, run commands, and automate repeatable work without leaving your terminal. A "terminal" is the plain text window programmers type commands into. In July 2026 OpenAI folded the separate Codex desktop app into the ChatGPT desktop app, so on a Mac or Windows PC it now lives as a tab next to normal chat.

Both are the four ingredients with the dials turned up. The model is the maker's current flagship. The tools are file reading, file editing, running programs, searching the web, and (through connectors, which lesson 4 covers) whatever else you plug in. The goal is whatever you type. The loop runs until the job is done, and on these tools "done" can mean hundreds of steps.

Where they run and what they cost

Both makers now ship the same agent on several surfaces, and which surface you pick changes the experience more than which company you pick. Tap through the comparison.

Tap a row to compare (as of September 2026)

Claude Code by Anthropic

An agent that reads a project's files, edits them, runs commands, and connects to your other tools. Anthropic's own one-line description.

Codex by OpenAI

OpenAI's coding agent: inspect code, make changes, run commands, and automate repeatable work, in their words. Since July 2026 its desktop app lives inside the ChatGPT desktop app.

Names, prices, and modes verified against Anthropic's and OpenAI's own pages on August 24, 2026. They change often. Check the maker's pricing page before you buy.

The pricing headline: neither is a separate purchase. Claude Code comes with paid Claude subscriptions (Pro at $20 a month, or $17 on the annual plan; Max from $100; Team and Enterprise seats). As of September 2026 the free Claude plan does not include it. Codex is included on every ChatGPT tier from Free up through Enterprise, with how much you can use it growing as you pay more (Go $8, Plus $20, Pro from $100). Both also run on pay-per-use API billing if you would rather not subscribe. The practical translation: if you pay for either chat app, try the agent before you spend another dollar.

Who they are for

Programmers, first and honestly. The default install is a text window, the vocabulary assumes you know what a repository is, and the examples in the docs are about writing tests and fixing bugs. If you hire a developer for your website, this is now the tool they reach for, and it is worth knowing that when you read their invoice.

But "built for programmers" and "only useful to programmers" are different claims, and the second one is false. Both products have grown surfaces aimed at people who never touch a terminal. Claude Code has a desktop app with a visual view of every change, a web version at claude.ai/code where you hand off a task and come back later, and it runs inside the Claude phone app. Codex lives inside the ChatGPT desktop app most people already have. The programmer's tool and the general tool are converging into the same thing, and the thing is "an agent that works on files."

What a non-coder can do with one

Think of it as a very capable temp who is only allowed to work inside one folder on your computer. You choose the folder. Inside it, they can read, create, rename, reorganize, and run things. Some jobs that fit that shape for a Southern Utah business:

A folder of PDFs into a spreadsheet. Drop a year of supplier invoices in a folder. Ask for one sheet with date, vendor, amount, and category. It reads each file and builds the sheet.
Photo cleanup. Four hundred job-site photos from a gutter crew, named IMG_4471 and up. Ask it to rename them by date and sort into folders by month.
Website text from notes. Give it a folder with your current site text and a notes file. Ask for a draft of a new services page that matches the voice of the existing pages.
A repeated report. "Every Monday, read the bookings export in this folder and write me a one-page summary." Claude Code's docs describe scheduled runs; Codex's cloud version can be kicked off from the web.

None of that is coding. All of it is "read files, write files, run things," which is what these agents are. The trick is giving them a folder that contains exactly what they need and nothing you would mind them touching.

The hard part, simply

The hard idea: you are not giving it a task, you are giving it your computer. Or at least the part of it you point it at. People who understand this treat these tools carefully. People who do not learn the hard way.

Analogy: hiring a contractor to redo one bathroom. You do not hand them the keys to the whole house and leave for a week. You show them the bathroom, you put the good china somewhere else, and you come by to look. The contractor is probably fine. The precautions are not an insult; they are how you sleep.

Worked example: you want the invoice spreadsheet from the list above. The careful way: make a new folder called "invoices-2025," copy (not move) the PDFs into it, open the agent in that folder, and leave the permission mode on the default that asks before changing files. Ask for the sheet. It will prompt before writing the new file; you say yes. It never saw your tax folder, your photos, or your email, because you never pointed it there. The careless way: open it in your home folder with the no-prompts mode on because a video said it was faster. It still probably builds the sheet. It could also, in trying to be helpful, "tidy up" something you cared about. Both makers document the no-prompts mode as something to run only inside a walled-off machine. That is not lawyer caution. That is the people who built it telling you where the edge is.

An honest take on risk

Three risks, in the order you are likely to meet them.

It does the wrong thing confidently. Same model weakness as every chatbot (AI Basics covered hallucination), now with hands. It might rename the photos by the wrong date field. Mitigation: work on copies, read what it did, keep the approval prompts on for anything that changes files.

It reads something that tells it what to do. The lesson 2 problem. A file in the folder, or a web page it fetched, contains instructions, and the model treats them as part of the job. Mitigation: permissions. A read-only agent that gets tricked can only tell you something wrong. An agent with send-email that gets tricked can send email.

It costs more than you expected. These tools run the big models for many steps, and the subscription tiers have usage limits measured in rolling windows. On an included plan you hit a wall and wait; on API billing you get a bill. Mitigation: start on the plan you already have, give it small jobs, and look at the usage page after the first week.

What is not on the list: the model "deciding" to do something. It has no goals except the one you typed. Every scary story about agents traces back to a goal it was given, a file it read, or a permission it was handed. Those are all yours to control.

Try this yourself

If you have a paid Claude or ChatGPT plan, you can try an agent this afternoon without installing anything in a terminal. Open Claude Code on the web at claude.ai/code, or open the ChatGPT desktop app and find the Codex tab. Then make a small, copied folder and give it this.

This folder has copies of some of my business documents. Before you change anything: list every file, tell me in one line each what it is, and tell me which three you would want to read first if I asked you to build a summary of my last quarter. Do not create, edit, or delete any files yet. Ask me before you do anything other than reading.

You just ran an agent in read-only mode on a fenced folder. That is the entire safe-start recipe. If you liked what it found, the next lesson explains the machinery that made the prompt-before-acting behavior possible, and lesson 5 turns this afternoon into a plan.

Next lesson4. Harnesses revisited

Last updated September 17, 2026