How do AI coding assistants work?

AI coding assistants work by feeding a large language model two things: your request and the relevant parts of your code. The model predicts the code that fits, and the tool shows it as a suggestion or a diff you can accept. More capable tools retrieve context from across your whole repository and run an edit-run-check loop to catch and fix their own mistakes.

What happens when you ask for code?

The tool first gathers context — the file you're in, the code around your cursor, and often pieces pulled from elsewhere in the project. It packages that context together with your instruction into a prompt and sends it to the model. The model generates code that statistically fits the patterns it learned during training, and the tool drops the result back into your editor as a completion or a proposed change for you to accept or reject.

How does it know about the rest of my project?

A model only sees what you put in front of it, and you can't fit an entire codebase into one prompt. So assistants index your files — breaking them into chunks and storing searchable representations — then retrieve the most relevant pieces for each request. This retrieval step is why a good assistant can reference a function defined in a file it was never explicitly shown.

Better retrieval is a big part of why one assistant feels sharper than another on the same code.

What is the agentic loop?

Simple assistants stop at a suggestion. Agentic ones keep working through a cycle:

1. Plan — break your request into steps.
2. Edit — write changes into one or more files.
3. Run — execute the tests, a build, or the app.
4. Check — read the output and see whether it worked.
5. Repeat — adjust and try again until it passes, or hand back to you.

That loop is what lets an assistant fix a failing test on its own instead of leaving a broken suggestion for you to sort out.