What is an AI coding agent, and how is it different?

An AI coding agent is a program that reads, edits and runs files in your project by itself. A chat assistant hands you code to paste; an agent makes the change, runs it, reads the error and tries again. That loop is the whole difference, and it decides what you can build, what it costs and where it goes wrong.

How is an agent different from a chat window?

Put them side by side on the same task — "the login page breaks on mobile".

The practical effect is not that the agent writes better code. On a single snippet, the quality is roughly the same — it is the same model underneath. The effect is that the ten steps around the code disappear: finding the file, understanding the surrounding code, making a matching change in three places, running the build, reading the error.

That is where the hours were. It is also why people who try an agent after a year of chat assistants describe it as a different category rather than a better version.

What can an agent do that a chat window cannot?

Four things, and each one is a different kind of work:

• Change several files at once, consistently. Renaming a concept across a project is tedious and error-prone by hand, and trivial for something that can see all of it.
• Read the error it just caused. Run, fail, read, fix. The loop is the point.
• Look things up in your actual project rather than guessing from the name. Fewer confident wrong answers about code it has not seen.
• Do a job while you do something else. This is the one that changes how a week feels.

What does it need access to?

More than people expect, and this is worth thinking about once rather than never.

An agent needs to read your project, write to it, and run commands. That means the same access you have: your source code, your configuration, and anything a command it runs can reach. If your project holds secrets in a file, the agent can read them. If a command can deploy to production, the agent can run it.

None of that is an argument against using one. It is an argument for three habits:

1. Everything in version control (GitHub is free), so any change can be undone with one command.
2. Secrets in environment variables, never in the code.
3. Read the diff before you accept it. Not every line forever — but especially at the start, and always for anything touching payments, authentication or data.

Where do agents go wrong?

Three failure modes account for most of it, and all three are measured rather than anecdotal.

The short version: generated code runs far more reliably than it is correct. It invents software libraries that do not exist, and it invents the same names repeatedly, which is a supply-chain problem rather than a typo. And it will follow instructions it finds inside files it reads, because it cannot reliably tell your instruction from text that is shaped like one.

I have written those up with the numbers and the sources in what AI coding agents get wrong, measured. If you plan to put anything an agent wrote in front of real users, read that one next.

Do you need a terminal?

No, but it splits the tools into two families that suit different people.

Terminal agents — Claude Code is the one I use — run alongside your project. Most capable, least hand-holding, and they assume you are comfortable with a command line.

Editor agents — Cursor — put the same capability inside a code editor. Gentler, and still expects you to be reading code.

Browser builders — Lovable, Bolt, v0, Replit, Base44, Emergent — take a description and produce a working app in a browser tab. No installation, no terminal. They are not the same category: they are excellent at getting to something that works and much weaker once you need a specific change in a specific place.

There is a full side-by-side of the first two in Claude Code vs Cursor vs chat, and of the builders in which AI app builder.

What does it cost?

One subscription is enough. The most common mistake is paying for an agent, an editor agent and two chat subscriptions at once — about €80 a month for four ways to ask the same question. There is more on that in what AI tools actually cost per month.

Should you start with an agent or a builder?

Honest answer: it depends on whether you can read code.

If you can read code but write it slowly, start with an agent. That is the largest single gain available to you right now, and it arrives in the first afternoon.

If you cannot read code at all, start with a browser builder. Not because agents are too hard to operate — they are not — but because you cannot supervise what you cannot read, and an unsupervised agent is how people end up publishing something with a security hole they never saw. Build two or three small things in a builder, learn to read what it produced, then move.

If you are not building software at all and came here to find out what the term means: an agent is only worth it if you have a project. Without one, a chat subscription does everything you need.

Where the word "agent" gets stretched

Worth knowing, because the label is on a lot of things now. Broadly it means software that takes a goal, decides its own steps, and acts — rather than answering one question at a time. Coding agents are the version where the actions are file edits and commands.

You will also see the word on tools that automate work outside code — MuleRun and Creao sit in that space. Same idea, different actions. If a product calls itself an agent, the useful question is always: what is it allowed to do, and who checks it afterwards?