What is an agent IDE?
An agent IDE is a development environment where AI agents write the code and the person reviews it. Instead of files and a cursor, its main objects are tasks, agents and diffs: you describe a task, an agent works on it in an isolated copy of the repository, and you approve, change or reject the result, usually as a pull request.
A classic IDE helps you type code. An AI code editor helps you type code faster. An agent IDE assumes you mostly won't type it.
How it differs from an AI code editor
| Classic IDE | AI code editor | Agent IDE | |
|---|---|---|---|
| Who writes the code | You | You, with suggestions and inline edits | Agents |
| Main screen | Files | Files plus a chat panel | Tasks and diffs |
| Unit of work | A file edit | A prompt | A task, ending in a pull request |
| Parallel work | One branch at a time | Usually one agent at a time | Many tasks, each isolated |
| Your job | Write | Write and accept suggestions | Assign, answer questions, review |
| Examples | VS Code, IntelliJ | Cursor, Devin Desktop | Zokie, Conductor |
What an agent IDE needs
- Isolation per task. Each task runs in its own git worktree and branch, so agents never overwrite each other's files. See running agents with git worktrees.
- A task board. When several agents run at once you need to see at a glance which are running, which are waiting for you, and which are ready to review.
- Diff-first review. The main view is what changed, not the file tree. Review is where the person adds the most value; see how to review AI-written code.
- Bring-your-own agents. Good agent IDEs run the agents you already use, such as Claude Code, Codex, Gemini CLI or Cursor CLI, instead of locking you into one model.
- A way to ask you. Agents get stuck or need permission. The IDE should surface those questions and let you answer without hunting for the right terminal.
- Cost control. Running many agents adds up. Picking a smaller model for small tasks helps; see small model or frontier model.
When you should use one
An agent IDE pays off when:
- you often have more than one change in flight, such as a bug fix, a small feature and a refactor;
- the tasks are well described, so an agent can finish them without constant steering;
- you are comfortable reviewing code more than writing it.
It pays off less when you are exploring an unfamiliar codebase by hand or doing delicate work line by line. Keep a normal editor for that; most people use both.
An example: Zokie
Zokie is the agent IDE we build. It is a desktop app for macOS and Windows, about 14 MB. It has three modules: Board for every task by status, Code where an agent works on a task in its own worktree while you watch the diff, and Studio, a local notebook for data work. It runs Claude Code, Codex, Gemini CLI or Cursor CLI on your own subscriptions, and each task ends as a pull request.
FAQ
Is an agent IDE the same as an AI code editor?
No. An AI code editor (like Cursor) is still built around you editing files, with AI helping. An agent IDE is built around agents doing whole tasks while you assign and review.
Do I still need a normal editor?
Usually yes, for manual edits and exploring code. An agent IDE replaces the part of your day spent typing routine changes, not every part.
Which agents work in an agent IDE?
It depends on the IDE. Zokie runs Claude Code, Codex, Gemini CLI and Cursor CLI. Conductor runs Claude Code, Codex and others.
Is agentic coding the same as vibe coding?
They overlap. Vibe coding usually means prompting and accepting whatever works. Agentic coding with an agent IDE adds structure: isolated branches, tests and a review step before anything merges.