Skip to main content
Vibe Code Forum

Article: Path 3: Use a coding agent in the terminal

Who it suits: people comfortable typing a few commands, who want the AI to handle bigger multi-step jobs. Also experienced developers who want to delegate.

What these tools are

Terminal agents run in a terminal window inside your project folder. You describe a task, and the agent reads files, plans, edits code, runs commands and tests, reads the results and keeps going until the task is done or it needs you. Examples include Claude Code, OpenAI's Codex CLI, Gemini CLI and the open-source Aider. Several also have editor extensions or desktop apps, so "terminal" is less strict than it sounds.

What makes them different from chat:

  • They act, not just answer: they can run your app, see the error and fix it.
  • They work across the whole project for longer, larger tasks.
  • They follow a project instruction file (each tool has its own name for it) with your rules.
  • They can connect to other tools through MCP. See MCP explained.

How to start

  1. Get comfortable with the terminal and Git. Agents change many files quickly; Git is your undo button.
  2. Install one agent following its official instructions, and sign in.
  3. Open a terminal in your project folder and start it.
  4. First task: "Read this project and tell me how it works and how to run it." Check it understood.
  5. Then give real tasks with clear checks: "Add a contact form that emails me. Done means the form shows a success message and the test passes."

Strengths

  • Handles tedious multi-file work: renames, upgrades, adding tests.
  • Can verify its own work by running commands.
  • Scriptable and repeatable for people who like control.

Watch out for

  • Permissions. Agents ask before running commands or editing files. Read what they want to run, especially anything that deletes, installs or pushes. Don't approve everything automatically on a project you care about.
  • Cost. Long agent sessions use a lot of tokens. Set limits and check usage. See Keep costs under control.
  • Drift. On long tasks, agents can wander off-plan. Ask for a plan first, approve it, and keep tasks small.

See also: Chat vs agent: what changes.

More in Pathways