An AI coding agent is a model that does more than answer in a chat window. It can use tools. Given a task, it reads the relevant files, makes edits, runs commands such as tests or a build, looks at the output, and decides what to do next. It repeats that loop until the task is done or it needs you. Claude Code, Cursor’s agent mode and similar tools work this way.
The difference from plain chat is that the agent can check its own work. It can run the test suite, see a failure and fix it without you copying the error back. That makes it good at multi-step jobs: adding a feature that touches several files, updating a dependency and fixing what breaks, or tracking down where a bug starts.
Agents act on your machine or your accounts, so permissions matter. Most tools ask before running commands or editing files, and let you choose what to allow automatically. Read those prompts. Be careful with anything that deletes files, pushes code, touches production or spends money.
A few habits make agents much more useful. Give them a way to know they are done, like a test that should pass or a page that should load. Work on a separate Git branch so you can review everything at the end. Keep tasks focused, one feature at a time. And review the final diff like you would a teammate’s pull request.
An agent that opens a pull request still needs someone to review it. That part of the process has not changed.
In the wild
Gave the agent the failing test and told it to stop when it passes. It changed three files. Reviewing the diff now.
