what is a coding agent?

the short answer

A coding agent is an AI program that doesn't just suggest code – it works: it reads and writes files, runs commands, tests its own output, and keeps iterating until the task is done. You state the goal – it figures out the steps.

The page you are reading right now was built this way. A human wrote the specification, an agent wrote the code – every step is public in the recipes.

chatbot → copilot → agent

Three levels worth telling apart, because everyday language calls all of them "AI":

1. Chatbot (e.g. the ChatGPT web interface): you ask, it answers. You copy the code out, paste it in, run it yourself. It cannot see your project.
2. Copilot (e.g. in-editor code completion): it suggests lines while you type. It speeds you up, but you still do the work, in small pieces.
3. Agent (e.g. Claude Code, Codex): it receives the task – "build a login page with a magic link" – and walks through it on its own: reads the existing code, writes the new code, runs the tests, fixes its own mistakes, and reports when it's done.

The difference is not one of degree but of kind: an agent runs in a feedback loop (it works in a cycle: acts, inspects the result, and decides the next step based on what it sees), which is what makes multi-step, real work possible.

how does an agent work?

A typical session looks like this:

1. You describe the goal in a prompt – in plain sentences, in your own language.
2. The agent explores the project: it reads the relevant files.
3. It makes a plan and starts executing: modifies files, runs commands (instructions executed in the terminal – e.g. starting the tests or building the project).
4. It checks its own work: runs the tests, reads the error messages.
5. If it finds a mistake, it fixes it and retries – this is the point: it doesn't stop at the first draft.
6. At the end it summarizes what it did, and you review it.

Your role doesn't disappear – it moves up. The what and the why stay yours; most of the how becomes the agent's. That role is what you learn on this platform.

why now?

Two things changed in 2025–2026. First, the models (Claude, the GPT family) got good enough to carry multi-step work through with error correction. Second, terminal-based agent tools appeared – Claude Code and Codex – that connect the model to the files and command line of your machine. Together they mean: one person who states goals clearly and reviews carefully can ship a working product alone.

This does not make knowing code useless – it means the entry barrier has moved: from syntax to systems thinking and quality control.

what will you learn here?

The curriculum starts from the very beginning, and stands on evidence throughout:

• preparing your machine and installing the two agents (Claude Code, Codex) – this module,
• your first working mini-build on your own machine,
• then module by module: writing specifications, design systems, databases, payments, going live – always through real, replayable working sessions.

The next lesson prepares your machine: we'll check what you need for all of this to run.