1. Coding agents

A coding agent is an LLM in a loop that can search a code repository, edit files, and run commands such as tests, then use their output to decide its next step. It is the basic agent loop you already know (Agents in LLM Engineering 101) with a repository and a shell as its environment. The basic principle is that code comes with its own grader. After each edit the agent can run the tests, the type checker, and the linter, read the errors, and try again, so it does not have to be right the first time.

That makes it the right tool for a bug fix or feature that tests can check, and for repetitive multi-file work such as a library migration or an API rename. It is the wrong tool for a one-line change you already know how to make. Type it yourself or accept an inline autocomplete. It is also the wrong tool for work with no automatic check, such as a visual tweak or an untested legacy module. There, have the agent propose a diff for a human to review instead of running an unattended loop.

This section is optional: companies that build coding tools ask about it in depth, and elsewhere it is a 5-10 minute follow-up. The question is usually "how does it find the right code, change it, and know the change works?" A mid-level answer explains that loop, picks one edit format, and writes a reproduction test first. A senior answer also defends a search combination, names how an agent passes weak tests without fixing anything, and structures a multi-hour task for review and rollback. The chapters follow the agent's own work order: find, edit, verify, then sustain.