What a Coding Agent Is
Ask a chatbot to fix a bug and it hands you code to paste, written without ever seeing your files. Ask a coding agent and it opens the files itself, makes the change, runs the tests and tells you what happened. The difference is not a cleverer model. It is what the model is allowed to do.
The rule
A coding agent is a model with hands: tools, a workspace and a goal.
- Model — the part that reads, reasons and decides the next move. In Claude Code, that is a Claude model.
- Tools — the hands: read a file, search the project, edit a file, run a terminal command.
- Workspace — the project folder the agent works in, with the same access your user account has.
- Goal — your request, which also tells it when to stop.
| Autocomplete | Chat assistant | Coding agent | |
|---|---|---|---|
| What it sees | Code near your cursor | Whatever you paste | Any project file it chooses to open |
| What it does | Suggests the next lines | Replies with text | Edits files and runs commands |
| Can it check itself? | No | No | Yes, if it has a check to run |
Claude Code is Anthropic's coding agent. You start it inside a project, describe a task in plain language, and it works through your files. By default it asks before it edits a file or runs a command, so every change still passes through you.
The shortcut
Chat answers; agents act. So brief an agent the way you would brief a colleague who has the code open: name the outcome and the check, not the keystrokes. "Fix the weekly discount so test_weekly_plan passes, then run pytest" beats "write me a discount function".
Look around
You report: "Weekly-plan customers are charged the one-day price." Claude searches the project for "weekly" and reads
pricing.py.
There are two opposite traps. One is treating the agent like a chatbot: pasting code into the prompt instead of letting it read the real files. The other is treating it like an oracle: accepting "done" without looking. The agent has hands, but you still supply the eyes.
Drills
Answer fast — trust the rule.
What makes Claude Code an agent rather than a chat assistant?
Which request gives Claude Code both a goal and a check?
When you run Claude Code in your terminal, where do its file edits and commands happen?
Which of these does a coding agent NOT give you?
Remember
- A coding agent is a model plus tools, a workspace and a goal.
- Shortcut: brief it with an outcome and a check, because it can act on both.
- Trap: pasting code as if it were a chatbot, or trusting "done" as if it were an oracle.