# Agentic coding tools: what to compare

> Claude Code, Codex, and Cursor now share the same core features. Here are the five differences that still matter when you pick an agentic coding tool.

By Antoine van der Lee. Last updated October 28, 2026.
Canonical URL: https://aifundamentals.dev/articles/agentic-coding-tools/

Agentic coding tools, also called agentic AI coding tools, are AI coding agents that read your code, edit files, run commands, and check the result without you approving every step. Claude Code, Codex, and Cursor are the three I use. Most comparisons rank them by features, but that list changes with every release.

The good news is that the three have grown towards each other. The core features are almost the same, so the choice depends on a few differences you can check yourself. I checked the documentation of all three on 28 October 2026. In this article, I show what the tools have in common, the five things that still differ, and how I would pick one.

## What do agentic coding tools have in common now?

Three features used to set tools apart. Today all three tools support them, according to their documentation for [Claude Code](https://code.claude.com/docs/en/memory), [Codex](https://learn.chatgpt.com/docs/agent-configuration/agents-md), and [Cursor](https://cursor.com/docs/rules):

| Feature | Claude Code | Codex | Cursor |
|---------|-------------|-------|--------|
| Reads `AGENTS.md` | Yes, when there is no `CLAUDE.md` | Yes | Yes |
| Subagents | Yes | Yes | Yes |
| Hooks that run your scripts | Yes | Yes | Yes |

The first row is the most useful one. Codex reads `AGENTS.md` before it does any work, and Cursor supports it in the project root and in subdirectories. Recent versions of Claude Code read it as well, but only when your repository has no `CLAUDE.md` in the working directory or above it. If you already have a `CLAUDE.md`, add one line at the top to import your shared file:

```markdown
@AGENTS.md

## Claude Code

Add Claude-specific instructions here.
```

This is why I keep my instructions in one `AGENTS.md`. It works across tools, so switching tools does not mean rewriting them. For how to write that file, see [how to write an AGENTS.md](/articles/how-to-write-agents-md/).

## Which five differences still matter?

When the features match, look at how each tool handles these five things.

1. **Where it runs.**\
   [Claude Code](https://code.claude.com/docs/en/permission-modes) runs in the terminal, in VS Code and JetBrains, and in a desktop app. Codex runs in the ChatGPT desktop app, the CLI, and the IDE extension, and it has cloud environments. Cursor is an editor, and it can hand work to cloud agents that run on their own VM and branch. Pick the place where you already spend your day.
2. **How permissions work.**\
   Claude Code has [permission modes](https://code.claude.com/docs/en/permission-modes), from asking before most actions to running everything, and deny rules that block an action in every mode. Codex splits it in two, as its [approvals and sandbox documentation](https://learn.chatgpt.com/docs/agent-approvals-security) explains: the sandbox sets what the agent can technically do, and the approval policy decides when it must ask. Check which model matches how much you trust an unattended run.
3. **How you control subagents.**\
   Claude Code [subagents](https://code.claude.com/docs/en/sub-agents) get their own prompt, tools, and permissions. Codex lets you define [custom agents](https://learn.chatgpt.com/docs/agent-configuration/subagents) with their own model configuration, but they inherit the sandbox and approvals of the parent session. In [Cursor](https://cursor.com/docs/subagents), subagents share the parent's checkout unless you ask for isolation, which gives each one its own Git worktree or cloud VM.
4. **Which models you can use, and what they cost.**\
   Subagents multiply usage. Codex's documentation says subagent workflows use more tokens than a single agent doing the same job. Check whether you can choose a model per subagent, and what a long session costs.
5. **How well it fits your validation loop.**\
   All three have hooks, but they differ in events and in where the hooks can run. A hook that runs your tests when the agent stops is only useful if it fires in the places where you use the tool. Codex, for example, asks you to review and trust a hook before it runs, and Cursor's cloud agents run only some of the hook types. See the hooks documentation for [Claude Code](https://code.claude.com/docs/en/hooks), [Codex](https://learn.chatgpt.com/docs/hooks), and [Cursor](https://cursor.com/docs/hooks).

## Why do subagent settings deserve extra attention?

Because the defaults can work against you. In my analysis of 1,674 sessions, 85% of subagents ran on my main model while they were only searching. You can read the numbers in my [token usage analysis on SwiftLee](https://www.avanderlee.com/ai-development/reduce-token-usage-claude-code-codex-cursor/).

Isolation is the other trap. When subagents edit files at the same time in one checkout, they can overwrite each other. I described the mess this caused in [subagent orchestration](/articles/subagent-orchestration-for-coding-agents/). Since then, I tell parallel subagents to use their own worktree.

## How should you choose?

My recommendation is to stop looking for the best tool and look for the best fit for your workflow:

- **Start from where you work.** If you live in the terminal, try a terminal-first tool. If you live in an editor, start there.
- **Test with a real task.** Give two tools the same small task in the same repository with the same `AGENTS.md`. Compare how much you had to correct, not how fast they finished.
- **Check the controls.** Look at permissions, model choice for subagents, and hooks before you commit to one.
- **Keep the valuable parts out of the tool.** Your `AGENTS.md`, your test scripts, and your hooks logic are yours. The more of your setup lives in the repository, the less it costs to switch later.

I would not pick a tool from a benchmark table. Models and features change every few months, but your instructions and validation loop stay the same. Those two decide most of the quality you get, which is why I wrote about [validation loops for agent-written code](/articles/validation-loops-for-agent-written-code/).

## Do you need only one tool?

No. I use more than one, because a tool that is great for a long background task is not always the one I want for a quick edit. If you do use several, the shared `AGENTS.md` and a few scripts in your repository keep them consistent. For the bigger picture, read [what agentic coding is](/articles/what-is-agentic-coding/).

## Conclusion

Claude Code, Codex, and Cursor all read `AGENTS.md`, support subagents, and run hooks. Choose between them on where they run, how permissions work, how you control subagents, what models cost, and how well they fit your validation loop. Keep your instructions and checks in the repository, so the tool stays replaceable.

To see how your current setup uses models and subagents, run the free [token cleanup prompt](/token-cleanup-prompt/). For the full approach, including validation, join the waitlist for [the AI Fundamentals course](/course/). Feel free to contact me or [tweet me on Twitter](https://twitter.com/twannl) if you have any additional tips or feedback.

Thanks!

## Frequently asked questions

### What are agentic coding tools?

Agentic coding tools are AI coding agents that read your code, edit files, run commands, and check the result without you approving every step. Claude Code, Codex, and Cursor are the best known.

### Which agentic coding tool is best?

None is best for everyone. Claude Code, Codex, and Cursor now share the same core features, so choose on where the tool runs, how permissions work, how you control subagents, what the models cost, and how well it fits your validation loop.

### Do Claude Code, Codex, and Cursor all read AGENTS.md?

Yes. Codex and Cursor read AGENTS.md. Claude Code reads it when there is no CLAUDE.md in the working directory or above it, and a CLAUDE.md can import it with an @AGENTS.md line.
