What a coding agent actually does.

A chatbot takes your question and answers it. A coding agent does something different: it reads your actual files, forms a plan, makes a change, then runs your tests to find out whether the change was any good. If the tests fail it goes round again.

That loop is the whole trick. The model is not smarter than the one in a chat window — it is just allowed to check its own work, which is the difference between a confident guess and something that compiles.

• Work you could do but would rather not. Renaming something in ninety files, writing the tests you kept putting off, moving a project from one library to another.
• Reading a codebase you have never seen. It will find where a thing is set far faster than you will.
• The first draft. Not the version you ship — the version you argue with.

It is confident when it is wrong. It cannot tell a requirement you forgot to mention from one that does not exist, so it will cheerfully build the wrong thing beautifully. And it has no stake in the outcome — it will not be the one explaining the outage.

Which is the same boundary we build into everything here: a decision the data supports should not wait for a person, and a decision that needs someone accountable should refuse to be made by software at all.

A coding agent is a very fast junior who has read everything and understood some of it, never gets bored, and needs reviewing. That is genuinely useful. It is not the thing the headlines describe, and anyone telling you it replaces the review is selling something.