You already know the basic interaction: describe what you want, read the AI’s answer, and decide what to do with it. A coding agent adds a workspace and tools to that conversation.
It can inspect the files in a project, make edits, run commands, and show you the result. That saves you from carrying every fragment between the chat and your editor. It also makes a vague request more consequential.
| In a chat | With a coding agent |
|---|---|
| You paste in the context | The agent inspects the project |
| The AI suggests an answer | The agent can edit the real files |
| You run the commands | The agent can run checks for you |
| You carry the answer out | The result already exists in the workspace |
Your job does not disappear
The agent can do the mechanical work. You still decide:
- What outcome matters.
- What it may change.
- Which decisions require your judgment.
- What evidence would make the result trustworthy.
That division is the foundation of this course. You are not learning a magic prompt. You are learning a small working loop: scope, plan, act, verify, review.
How the agent acts
Most coding agents work through the same interfaces programmers and scripts already use:
- A CLI is a program the agent runs as a terminal command. It is often the clearest route for files, Git, builds, tests, and local tools.
- An API lets code send structured requests to a service. The agent can use it directly or through a small command that handles details such as authentication and pagination.
- MCP gives an agent host a standard way to discover tools and live context. It is useful for shared service connections, but it is not a replacement for every CLI or API.
You do not need to operate all of these yourself. The important habit is to prefer a clear, inspectable tool surface over repeated clicking or improvised steps. The field manual has a fuller decision guide in Choose the right tool surface.
The safest first move is often read-only: ask the agent to explain the project before asking it to change the project.