Kuala Lumpur
← All use cases

中文

Building software · Claude Code + GPT

Run two coding agents side by side, one on frontend and one on backend

Idea + starter prompt, not run by us yet

For: Solo builders and small teams who juggle more than one coding agent

A poster put Claude Code on the frontend and GPT's Codex CLI on the backend, and ran both for 15 days in a shared workspace called Tutti. Both could see the same files and task outputs. They passed task cards between agents instead of re-explaining, and used @ references to pull earlier Claude chats into the other agent.

Steps

  1. Decide the split first: which folders belong to the frontend agent and which to the backend agent.
  2. Put both in a shared workspace or a shared project folder, so they read the same files.
  3. Pass work with a short task card: goal, files involved, what is done and what is left.
  4. Keep a simple board of tasks marked in progress, done, failed or stopped, so you can see where things stand.
  5. When a task fails, check the last step it ran before restarting. That was the quickest recovery in the source.
  6. Give it more than two weeks before you trust the setup. The poster says 15 days was too short for a hard test.

Starter prompt

Starter prompt. Change the [brackets] first.

I am building [project] with two coding agents. One handles the frontend in [frontend folder] and the other handles the backend in [backend folder]. You are the [frontend or backend] agent.

Trial first. Do not start on the real project. Make a scratch repo in [trial folder] with one toy endpoint and one toy page, and run one full task-card round there. Nothing from the trial goes near my real repo.

Before you change anything in the real project, write a short task card for the other agent: the goal, the files you will touch, the API or data shape you expect from them, and what you need back. Show it to me and wait for my OK.

Work only inside your own folder, on a git branch. If you need a change on the other side, write it as a new task card and do not edit their files. After each task, list what you changed and what is still open.

Stop and wait for my yes before you commit, push, merge or deploy. Showing me the diff is not permission to do any of these.

Original post: Reddit · found via agentusecases.ai. Written up in our own words. We haven't run it, so test it on made-up data first.