What is a coding agent?

The category that Claude Code, Codex, and Cursor all belong to. Not an AI that talks about code. One that reads your files, edits them, runs them, and checks its own work.

Last month I asked one to migrate a database table, write the tests, run them, fix the two that broke, and open a pull request for my engineer to review. I approved it from my phone over coffee. I never opened my laptop. Two years ago that was a junior engineer's afternoon.

The thing that did the work was a coding agent. It is the most important tool category a CEO can understand right now, and most never learn the word for it.

What it is

A coding agent is an AI model (a large language model) wrapped in enough tooling to act on a codebase instead of just describing one. The chat version of AI hands you text. You copy it, paste it into your editor, run it, hit an error, and paste the error back. You are the hands; the model is the mouth. A coding agent closes that loop. It reads the files, makes the edit, runs the command, reads the output, and tries again until the thing works.

If you have read What is a harness?, you already have the frame. A coding agent is a harness pointed at code. The model is the brain. The harness is the hands. The codebase is the workshop. Strip any one of those out and you are back to a chatbot.

The names you will hear: Claude Code (Anthropic), Codex (OpenAI), Cursor, and Gemini's command-line tool. Same category, different vendors, and they leapfrog each other roughly every month. The open harnesses operators run, OpenCLAW and Hermes, are built on the same primitive.

Why CEOs care

You may never write a line of code. Your engineers already run one of these every day, and the tool they pick shapes how fast your product ships. That alone is worth understanding.

The part nobody tells you: the category outgrew code. A coding agent is, underneath, an agent that can read files, run commands, and check results. Point it at your ops scripts, your exported data, your docs folder, and it does operator work, not just engineering work. I run one to pull sales numbers, tag support tickets, and draft weekly reports. The "coding" in the name is a historical accident, not a fence.

The other reason: the vendor of the month is noise. The category is the signal. A CEO who learns what a coding agent is can switch tools in an afternoon when the lead changes hands. A CEO who learned one product's buttons has to start over. Learn the category, rent the vendor.

How it differs from the chatbot you already use

The chatbot in your browser advises. It is a brilliant intern who can only talk to you through a window and never touches the work.

A coding agent has hands. It runs in your actual environment, with permission to open files, run commands, and act on what it finds. You stop being the copy-paste layer between the model and the result. That single difference is what turns AI from a faster Google into something that gets work done while you are asleep.

Where you'll see it

• In every technical hire's setup, where they are choosing between Claude Code, Codex, and Cursor right now.
• In nearly every workflow on this site. The Desk Theory stack runs on a coding agent.
• In the "Codex vs Claude Code" debate your team is probably having on Slack.
• In every vibe coding story where someone non-technical ships an app over a weekend. The coding agent is the thing doing the building.
• In the operator harnesses OpenCLAW and Hermes, which dress a coding agent up for running a business, not just a repo.

What to do next

If you have never watched one work, that is the thing to fix this week. Start with What is OpenCLAW? for the harness most operators run, or What is Codex? for OpenAI's version of the same idea. Then sit and watch it run one real task end to end. It changes how you think about your week. Tell me what it built for you.