Skip to main content
Vibe Code Forum

Article: Chat vs agent: what changes when the AI can run commands

AI coding tools come in two broad modes, and many tools offer both.

Chat (or "ask") mode

You ask; the AI answers with text and code. You decide what to do with it: copy it, apply it, ignore it. Nothing happens to your project unless you make it happen.

Good for: explanations, small changes, reviewing code, planning, learning.

Agent mode

You give a goal; the AI works towards it by itself in a loop:

  1. Reads files to understand the project.
  2. Makes a plan.
  3. Edits files.
  4. Runs commands: installs packages, starts the app, runs tests.
  5. Reads the results (errors, test failures) and tries again.
  6. Stops when it thinks it's done, or asks you a question.

Good for: multi-file features, upgrades, fixing a failing build, writing tests, repetitive changes across a project.

What changes for you

You become the reviewer. The agent can do in minutes what used to take hours, and it can also make a large mess in minutes. Your job shifts from typing code to setting clear goals and checking results.

Clear "done" matters more. "Add a search box" leaves the agent to guess. "Add a search box above the list that filters by title as you type; done when the existing tests pass and searching 'milk' shows only items containing milk" gives it something to check.

Permissions matter. Agents ask before running commands or editing files. Read what they want to run. Be very careful approving anything that deletes, force-pushes, touches production or installs unfamiliar packages.

Git is essential. Commit before starting an agent task, so you can see exactly what it changed and undo it in one step. See Git and GitHub explained.

A good agent routine

  1. Commit your current working state.
  2. Ask for a plan first, read it, correct it.
  3. Let it work, approving commands you understand.
  4. Review the changed files and run the app yourself.
  5. Commit if good; roll back if not, and try again with a clearer goal.

See Path 3: Terminal agents and Path 2: AI code editor.

More in Concepts