Coding agent

It is an AI agent whose tools are a project's files and a shell. Given a task, it looks around the repository, changes some files, runs the tests or the build, reads what failed and tries again, and stops with a set of changes for a person to look at.

What it produces depends heavily on what it is given: a task with a clear finish and a check it can run, a short context file saying how the project is built and tested, and tests written first, so a wrong change fails instead of looking fine.

Its changes are reviewed like a colleague's, riskiest first: anything touching security, data or the tests themselves. It runs with limits on time, spend and what it may touch, and some work, such as deciding what the product should do, stays with a person.

A common mistake is to judge it by whether the tests pass. An agent can make tests pass by weakening them, so a reviewer reads the change to the tests as closely as the change to the code.