The Nerd Uncle › Agents

What a coding agent actually does.

A chatbot answers you. An agent keeps working after you have gone. The difference is one loop, and it is the whole trick.

29 OCTOBER 2026 · 3 min read

Code on a screen, mid-edit. — The agent reads files like these, changes them, then runs the tests to find out whether it helped. Photograph: MikeRun (CC BY-SA 4.0)

Ask a chatbot something and it answers. The exchange is over.

An agent is given a job instead of a question. It does not answer — it goes away and works.

Claude Code, Codex and the rest all run the same loop. The branding differs more than the method does.

First it reads. Not a summary of your project: the actual files, as they are on disk right now.

Then it changes something. A line, a file, ninety files. That part looks impressive and is the least interesting thing it does.

Then it runs your tests — and this is the bit that matters. The machine finds out whether it helped, without asking you.

A chatbot that gets it wrong wastes thirty seconds of your time. An agent that gets it wrong finds out by itself, and tries again.

It stops when the tests pass. Not when it feels finished — when the project says so.

The honest summary

It is not smarter than the chatbot in your browser. It is the same kind of model, given three things a chatbot is not: your files, permission to change them, and a way to check its own work.

That is the entire difference. Everything else is detail.

Give it work that has a right answer — tests that pass, code that compiles, a build that goes green. Work where nothing can tell it that it failed is work you will have to check yourself, and then you have not saved anything.