S1 · E3 · Beginner · Season 1: AI Agents & Tools

Which AI coding agent should I use?

Explained with Pizza

A coding agent is an AI that works on your code by itself in a loop: it reads your files, edits the code, runs a command, and checks the result, then repeats until the task is done.

In short

  • Claude Code (Anthropic), Codex CLI (OpenAI), and Antigravity CLI (Google) all run the same loop: read, edit, run, check.
  • Cost, checked Nov 20, 2026: Claude Code needs a paid Claude plan (Pro: $20/month billed monthly, $17/month billed yearly) or an API key. Codex CLI is open source; OpenAI recommends using it with ChatGPT Plus or a higher plan, and you can also use an API key. Antigravity CLI has a $0 plan with a weekly agent quota.
  • Newer versions of Claude Code and Antigravity CLI start in auto mode. Check the mode before you begin.
  • Start with an ask-first mode; save full auto for a sandbox.

What a coding agent does

A chatbot gives you code to copy. A coding agent works inside your project. It runs in your terminal, the text window where developers type commands, and it can open files, change them, and run programs on its own.

Under the hood, all three agents in this lesson run the same loop: read the files, edit the code, run a command (for example, the tests), and check the result. If something fails, they repeat the loop.

Think of a pizza kitchen

A chatbot hands you a recipe. A coding agent walks into your kitchen, makes the pizza, tastes it, and fixes it until it is right.

An ask-first mode is a cook who asks before turning on the oven. A sandbox is a separate test kitchen: if something burns there, your real kitchen is fine.

The three agents

In June 2026 Google moved free users from Gemini CLI to Antigravity CLI.

Claude CodeCodex CLIAntigravity CLI
Made byAnthropicOpenAIGoogle
Open sourceNoYes (Apache license)No
Cost (Nov 20, 2026)Paid Claude plan (Pro: $20/month, or $17/month billed yearly) or an API keyOpenAI recommends ChatGPT Plus or a higher plan; you can also use an API key$0 plan with a weekly agent quota
Asking firstManual mode asks before most edits, commands, and network accessApproval settings decide when it asks; commands run in a sandboxDefault mode asks before edits and commands

Safety: who decides what the agent may do?

An agent that can edit files and run commands can also break things: delete a folder, install something, or push code you did not check. So each tool has two kinds of protection: asking you first, and limiting what it can touch.

Asking first: Claude Code's Manual mode, Codex CLI's approval settings, and Antigravity CLI's Default mode. Limiting: Codex CLI runs commands in a sandbox by default, a closed box that limits which files and network they can reach. A sandbox does not ask you anything; it only sets the walls.

Careful: newer versions of Claude Code and Antigravity CLI start in auto mode, so check the mode before you begin.

So which one?

  • Want a $0 plan? Antigravity CLI.
  • Want open source? Codex CLI.
  • Already paying for Claude or ChatGPT? Try the agent that comes with your plan.

One rule for all three

Start in the mode that asks first. Allow more, step by step, as you learn what the agent does. Use full auto only in a sandbox. And test the agent on your own project, because the best agent is the one that works well on your code.

Prices and plans change often. These were checked on November 20, 2026; look at each tool's pricing page before you decide.

Test yourself

5 quick questions. Pick an answer to check it.

1. What loop do all three coding agents run?
2. Which of the three is open source (Apache license), as of Nov 20, 2026?
3. You are new to coding agents. Which setting should you start with?
4. What does Codex CLI's sandbox do?
5. Why should you check the mode before you start Claude Code or Antigravity CLI?

Sources

Checked November 20, 2026.