Both Big AI Labs Built Tools for Programmers. "Coding Agents". Now They Are for You.

Anthropic, big AI company, took the tool they built for programmers, put a chat window on it, and handed it to everybody else. They called it Cowork. Then they published the logs.

Not a survey, but their own data, and 1.2 million sessions were sampled across more than 600,000 organizations.

Software development was only 8.7% of it.

Business process and operations were 33.4%. Content and copywriting was another 16.4%. The rest is stuff like: Pulling scattered updates into one report. Building onboarding checklists. Reconciling spreadsheets.

Most of what it does is work nobody thinks of as 'coding agent' work.

Sidebar... If anyone has ever wondered why I use single quotes instead of double quotes sometimes... I am aware it's improper... but I use them when writing code in Python, and it's a habit I can't break. I'm working on it. Also, I write these for me, so I remember this stuff. Not you... But thanks for reading.

Anyways, Cowork is not some separate thing Anthropic built from scratch for office workers. It runs on Claude Code as an engine. Claude Code is a coding agent. We will get to defining stuff in a minute. Anthropic's own words: they built Cowork to "extend the agentic capabilities of Claude Code" into a chat window. Boris Cherny, the man who created Claude Code, says Claude Code wrote most of it.

Then on July 9th, five days ago, OpenAI, another big AI company, did the same thing from the other direction. They folded Codex, their coding agent, into a new ChatGPT desktop app that puts Chat, Work, and Codex in one place, and they launched something called 'ChatGPT Work' on top of it.

Sam Altman's words, lowercase and all: "codex is the core of our new work product and what makes it so good."

Two companies. Different but similar philosophies. Same move, same month. Both of them quietly took the tool they built for engineers and pointed it at YOUR job.

But before we get into which company did what, we need to talk about what they actually built.

So here is Classroom Edition #3. By the end of this you will know what a coding agent actually is, where it fits in with all the AI vocabulary we have already covered in previous editions, what is inside one, exactly how Claude Code and Codex differ and which one is right for you, what happens when they click around your screen, why the name 'coding agent' is doing real damage right now, and enough info to start using one today.

Read the whole thing and you will understand these tools better than most of the people currently using them.

Quick note before we start: product plans, access, and feature names are accurate as of July 9th, 2026. Both companies change this stuff constantly...almost daily.

We have done two of these already, so let me stack them for you.

Classroom Edition #1 was the big map. Artificial intelligence is the giant umbrella term. Inside it, machine learning finds patterns instead of following specific rules. Inside that, deep learning unlocked the current wave. And a large language model, an LLM, is one type of model that came out of it.

These LLMs are just math. They predict the next token. The math is complicated but the concept is simple. The tokens turn into words so inferior humans, like me, can understand.

Classroom Edition #2 was the next ring down. Agentic AI is AI that takes actions toward a goal instead of just answering you. An AI agent is a thing you hand a goal to, and it runs a loop until the job is done. A chatbot talks. An agent does.

Today we go a little deeper.

A coding agent is an agent that can write and run actual software or code, look at cause and effect, locate bugs or errors, and fix them. It can do this all on its own. In a loop.

That same loop can sort a folder, reconcile a spreadsheet, pull scattered updates into one report, or run really boring workflows every week.

So the ladder is: AI, then machine learning, then deeper learning models, then LLMs, then agents, then coding agents. Each one a smaller box inside the one above it.

In the last couple of weeks, these 'coding agents' have become extremely powerful, in a way that matters to everyone... Not just technical folks.

You need to have a basic understanding of 'coding agents' regardless of your role or field.

... in my humble opinion.

'Coding agent' is the worst name in AI right now.

It tells millions of people that this thing is not for them. And it comes with a matching mental image... a black terminal window, green text, someone typing weird commands.

That image is doing real damage. It's the reason most people haven't opened one of these yet. The name told them not to.

To be fair, Claude Code does have that old command-line vibe. But Codex looks exactly like the ChatGPT app most people already use every day. Same idea underneath either one: build things and automate almost anything through plain conversation. That's what they were built to do.

Anthropic and OpenAI both set out to build an agent that could write software using a language model. Claude Code and Codex. To do that, they had to give it powers that have nothing specifically to do with writing software.

It had to read and write files on a real computer. Run commands. See what happened. Notice failures. Try again. Keep track of what it had already tried. Ask permission before doing something dangerous.

It's no longer a code-writing tool. It's a system that can complete complex multi-step jobs on a computer and work through them.

The work still needs to be bounded. It helps tremendously if you can explain what “done” looks like and how to check whether it did the job correctly. But the job itself does not need to involve code.

It can sort a folder. Reconcile a spreadsheet. Pull scattered updates into one report. Build an onboarding checklist. Run the same boring workflow every day.

Writing software is just the hardest, most checkable version of that kind of work. And that was the original intent.

That is what the 8.7% number is telling you. The engineers built the complex engine. But now anyone can drive the car.

So for the rest of this article, when I mention 'coding agent', I mean an agent that can operate a computer, which they happened to teach to write code.

That is why 'coding agent' is a horrible name now.

At the beginning, it was accurate. Now it is keeping a lot of people from realizing what these things can actually do... and who can use them.

Here we will go into some of the technical stuff surrounding these coding agents. You don't really have to know all of this stuff, or even understand what it means. You can start building things today with almost no knowledge. You can download the Codex app on your laptop and just ask it to do stuff... But you will have to eventually know what they are so I will mention them...

Last issue I mentioned harnesses so I will keep it brief. Everything that makes these tools useful, and everything that makes them potentially dangerous, lives in the harness.

The harness is the environment that wraps the model and does the actual work. It makes a LLM an agent. It runs the agent loop (will explain that again in a minute), hands the model its tools, and decides what it is allowed to touch.

The model is the brain.

The harness is the body it drives around in.

Claude Code is a harness. Codex is a harness.

A look inside the harness....

There are seven pieces worth knowing. The first three explain how it works. The last 4 explain how to make it useful without trusting it blindly.

Gather context, take action, verify results. Then do it again. And again. Those are Claude Code's own three words for it, straight out of their documentation. A chatbot answers once and stops. An agent runs laps until it is done or stuck, chaining dozens of actions together on the way. If you want more detail, see my last edition.

A tool is anything the agent can reach out and use. Read a file. Write a file. Run a command. Search the web. Run a test. This is how models escape the chat box.

And here is the part people skip... if you want an agent to fix your spreadsheet, it has to be able to open your spreadsheet. Duh.

There is no version of this where it helps you without being able to touch your stuff. Every single risk tied to codinhg agents flows from that one sentence.

To do cool things... you must be vulnerable. That's true in life also.

This is probably new to you, and it is the most useful thing in this entire article, so slow down here.

Claude Code has a throttle to assist in preventing devastating mistakes. While working you can hit Shift and Tab, and it will cycle through four options. Codex has similar options.

Plan mode: it can look, but it cannot touch. It reads everything, tells you exactly what it intends to do before it does anything, and it does nothing until you approve. Most users start here to plan then switch to auto.

Manual mode: it asks you before every single action. In other words... It slows you down tremendously.

Accept-edits mode: it automatically approves all file edits and basic filesystem commands without a safety check, but prompts you for risky commands. This is not a good idea. Don't do it. It will overwrite stuff you don't want it to.

Auto mode: it mostly just goes. There are still guardrails and approvals, but this is not a magic setting that makes every action safe. But most users use this setting to build or automate things. You can customize the guardrails.

(There are 6 modes under the hood. And you can customize approve/no approve actions further if you want. But the main 4 are the only ones a new user needs to think about.)

If you take one thing from this issue, take this: start in plan mode. Every horror story happened when someone skipped plan mode.

The context window is everything the agent can currently "see" at once. Your instructions, the files it has read, what it already tried. It is finite. When it fills up, quality falls off a cliff. (The quality actually gets worse before the window fills but that's a topic for another day and it is called context rot.)

I have written before that a model which was sharp at the start of a conversation seems almost dumb by the end of it. That is the context window filling up. It is still the thing that bites people hardest, and it is why these tools now have to summarize themselves mid-task just to keep going.

The more information you feed a model in a single session, the dumber it gets. And they tend to forget the stuff in the middle the most. They will remember the beginning and end of a conversation.

Now this is the part that many people don't use at all or don't use appropriately.

You can drop a plain English text file into your 'project folder' that tells the agent how you want it to work. What this project is. How to check its own work. Everything that is important for the project. What to never touch, etc.

Claude Code calls it CLAUDE.md. Codex calls it AGENTS.md. There is no code in it. There is no format to learn. You just write, in normal sentences, what you want.

An example of one of my AGENTS.md files below.... Just text with high level instructions.

You can tell your agent to open every response with 'I like meatballs' and it will... everytime you talk to it.

There are definitely best practice methods for using these files, and how long they should be, but again... topic for another day.

A study this year ran OpenAI's Codex over 124 real code changes, once with an instructions file and once without. Same quality of work, but 28.6% faster on about 10% fewer tokens. Again, think of a token as just a chunk of a word, and it is the unit these things get billed in, so fewer tokens means a smaller bill. (If you have a Claude or ChatGPT subscription you already have tokens in the form of a monthly allowance.)

The instructions file does not make the agent smarter. Stops it from wandering around your project guessing, and wasting time and money.

My 'second brain', that I go on and on about, has one master instruction file, to ensure models can retrieve all of my information and research efficiently.

Make sense? Of course it does. You would do exactly the same thing for a new hire.

A skill is a folder with a text file in it. That's all a skill is.

You write down how to do a task once, in English, and the agent can pick it up whenever that task comes around again. The clearer and more specific you make it, the better chance it has of knowing when to use it.

Anthropic published the format as an open standard in December. OpenAI switched it on in Codex the next day, and got rid of their own competing feature about a month later. By OpenAI's own count, the share of active Codex users reaching for a skill went from 5.4% on March 1st to 26.6% on June 11th.

Two companies actively trying to eat each other, and one of them shipped support for the other's file format in 24 hours. They had been quietly building toward it for weeks, which if anything makes the point better. That should tell you how much this layer matters.

You can see a few skills I made public here. Click one of the 3 skill folders > click on the SKILL.md. It's just text. That's a research skill I published that will allow you to say something like...

and you will get a quality in-depth dossier with checks and balances about bananas and their relationship to Italy.

My agents make most of these skills for me. I have agents create them anytime I find myself doing repetitive tasks. Here is a skill called 'brain-advisor' that combs through all of my data and offers suggestions, as an advisor.

More text.

No need to know code to use a coding agent.

Remember this...

Writing "never delete my files" in your instructions file is just a request. And models make mistakes frequently. The model can, and will, ignore requests from time to time. It's just text.

Hooks are different. A hook is a piece of actual code that runs before the agent is allowed to act, and it can say no. That is not a request. That is enforcement.

For example: block any command that deletes files outside one specific folder. The agent can ask all it wants. The hook still says no.

You don't have to know code to build a hook. You can ask the agent to build them.

Anthropic says this in their own documentation, plainly, which I respect them for. Requests are polite. Hooks are load-bearing. I have spent a lot of time building hooks and validation loops into my own setup for exactly this reason.

Again... you don't need to know any of this to start using coding agents today.

There are others, but these are the 2 that matter. We will start with Claude.

Claude Code is Anthropic's agent. It's meant to work next to you, on your workstation.

The original version lived in the terminal. You open a real folder on your computer, run claude, and talk to it while it reads files, writes code, and runs commands. It can also be used through editors, a desktop app, and the web now. But the feeling is still the same.

You are in the cockpit beside it.

You can interrupt it. Ask why it chose something. Tell it to stop. Point at the wrong turn. Make it explain the plan before it touches anything.

That is Anthropic's strategy... give people a powerful agent close to their real work, then give them lots of custom ways to steer, constrain, and inspect it.

The Claude models are especially useful when the work needs a lot of reading, reasoning, and back and forth. It's great for understanding an unfamiliar project, planning a complicated change, investigating why something broke, or turning a messy idea into a concrete build plan.

Below is an image of Claude Code creating a plan about how to best explain 'Plan Mode'. Its plan was better than mine... but I used mine.

So Claude Code is a very good fit if a job is fuzzy at the start and you want to stay involved as it becomes clear.

It's a bit pricier than ChatGPT's frontier models.

Claude Code is a good starting point for:

Developers working in a real codebase every day.

Operators and analysts who want to build internal tools and want to see every step.

People who have a project folder, a clear outcome, and the patience to steer.

Anyone who wants to learn how agents actually work instead of watching a black box make decisions.

It is less ideal for someone who has never opened a terminal, has no project folder, and just wants help with normal computer work today.

Anthropic knows this. That is why it has also built friendlier surfaces around the same core engine, including Claude Code on the web and Cowork.

But its main attraction is still Claude Code.

OpenAI took a different path.

Codex also has a CLI. You can open a folder, run codex, and let it inspect files, make edits, run commands, and show you a diff.

A diff is just a before-and-after view of what changed. Additions show up as additions. Deletes show up as deletes. It is the receipt you review before trusting the work.

But OpenAI is pointing Codex at normal and technical work. It's way easier for non-technical workers to master.

The new ChatGPT desktop app combines Chat, Work, and Codex.

Chat is for asking questions.

Work is for projects like documents, research, spreadsheets, slides, and files.

Codex is for harder technical work: code, folders, commands, repositories, and more involved tasks.

That is OpenAI's strategy: start with the ChatGPT app people already understand, then let the agent take on more of the job when the task requires it.

It is a better door for beginners.

You do not have to wake up one morning and decide you are a terminal person. You can start in a familiar app, hand it a real task, and only go deeper when you need to.

Codex now has Computer Use inside the ChatGPT desktop app. I used it... it's awesome.

That means it can see and operate an app you approve, or your entire computer. And you can tell it what to do... then watch it. It can click, type, and work through a visual task like a person would.

This is useful when the work happens in a spreadsheet, browser, old company software, or an app with no clean connection for the agent to use.

And it is why Codex is the better first agent for most beginners right now.

You can give it a normal task in a normal app, instead of starting with a repository and a terminal.

But Computer Use is not the best way to do everything. It is the broadest way. It is usually slower and less reliable than a proper integration, a file, or a command.

For a first task, approve one app. Give it one small job. Watch it.

I asked Codex to make some changes to my newsletter site the other day with very vague instructions, and watched it make creative choices on my own laptop for me. Really cool! See video below...

Codex is a good starting point for:

Beginners who already use ChatGPT and want to get actual work done.

People with a defined project they can check: organize a folder, clean a spreadsheet, turn notes into a report, or build a small tool.

Technical people who want to hand off a contained coding task, review the diff, and move on.

Teams that want separate, reviewable tasks running in cloud sandboxes instead of every agent touching one live machine.

SO WHICH ONE SHOULD YOU USE?

For most beginners: start with Codex in the ChatGPT desktop app.

Remember that black terminal window I said scares people off? On the right, that's it. Claude Code CLI. On the left is the new ChatGPT/Codex app. These do basically the same thing. One of them is just far more inviting.

Codex has the easiest front door. It keeps Chat, Work, and Codex in one place. And when you truly need an agent to use a normal application, Computer Use is there without having to build your own weird little laboratory first.

Start with Claude Code if you want to learn the deeper version of this quickly, or you are already working in project folders and want an agent sitting beside you while you think.

Then there is the boring but correct answer.

If you already pay for Claude or ChatGPT, start with the one you already have.

The difference between using Claude Code and Codex matters a little.

The difference between using one of them and using neither matters a lot.

Do not start by building an app.

That is what every demo wants you to do. The demos are made by people who have been building apps daily. They are not really the normal first experience.

Start with one small job on something you already understand well.

A copied folder of documents. A spreadsheet with obvious cleanup problems. A pile of notes that needs to become a clean outline. A document library that needs a searchable index. Something like that.

The main rule is simple: you must be able to recognize a bad solution if it gives you one.

If the agent sorted your receipts incorrectly, you would notice.

I'm not going to go into a full demo of either. Those are available everywhere. The important thing is that you start.

You do not need to become a programmer to use an agent well.

You need to know what you want. You need to define boundaries. And you need to be able to look at what came back and say, "No. That is wrong. Try again."

That is not a new skill. You have been doing it your entire career.

The tools are getting better at the work.

All positions are transitioning to 'Supervisor of Machines'.

That is the assumption of Claude Code and Codex. They wouldn't be building these systems the way they are if that wasn't true.

Anthropic is building the cockpit for people who want to steer a powerful agent closely. OpenAI is building the work app that makes the same kind of agent easier for everybody else to pick up and use.

Different but similar. Humans are still involved.

Start small. Keep a copy. Ask for a plan. Check the receipts.

Download ChatGPT/Codex or Claude Code on your laptop today. Ask it a simple question.

Next Classroom Edition we go into hallucinations properly. Everything in today's article gets significantly more dangerous when the agents are confidently wrong.

Every technical term in this article is defined in AI Homeroom, which is free, and Barnabas built in 10 minutes:

https://dustincole-data.github.io/ai-homeroom

You now understand coding agents better than most of the people building on top of them.

Thanks for reading 🙏

🔗 LinkedIn: linkedin.com/in/dustinwcole

🌐 Site: dustincoledata.com

📩 Reply with what you want explained next. I read everything.

— Dustin