Coding Agent.

Early AI coding aids completed single lines, much like word suggestions on a phone. A coding agent goes further. It receives a task such as 'Add a phone field to the contact form and store it in the database', finds the affected files itself, edits them, runs the tests, reads error messages and improves its code until the task is done. Technically it is a large language model with tools, see tool calling: reading files, writing files, running commands in the terminal.

Coding agents run in different environments. Some work in the terminal directly on the developer's machine, others inside the IDE, others again in an isolated cloud environment where they submit changes as a pull request on GitHub. What they need to know about a project they read from project files such as CLAUDE.md and AGENTS.md, recurring procedures from agent skills. Running several agents in parallel is possible with Git worktrees without them overwriting each other's files.

The benefit is real but not magic. Agents are strong at clearly defined tasks in well-structured projects: fixing bugs, writing tests, adding features following a pattern, refactoring. They are weaker with unclear requirements, architecture decisions and anything needing domain knowledge outside the code. They can hallucinate, invent libraries or adjust tests so they pass instead of fixing the bug. Hence the rule: every change is read by a human, and automated tests such as end-to-end tests are mandatory.

An important distinction concerns vibe coding. In vibe coding you let the AI get on with it without understanding the code in detail, often fine for prototypes. Professional use of coding agents means clear tasks, project rules, code review, tests and version control with Git. Security matters too, because an agent with terminal access can also cause damage. It should work only with the permissions it needs and never have access to production secrets it does not need to know.