Real story · Website production

Claude Code and Codex built one website together

They built the same site, caught different problems, and checked each other's tests.

Claude Code and Codex agents connected to each other while building and reviewing a live website in October
Build this site together. Split the work and review each other before shipping.

Why this gets messy

One coding agent can build a site fast, but it can also miss its own mistakes.

Claude Code and Codex make different choices and notice different problems. We wanted both of them on the same job so they could challenge each other.

Put the whole job in one place

I put Claude Code, Codex, the repository, and the live site on one October canvas. The agents stayed separate, but they could message each other, hand off tasks, and inspect the same result.

Give each agent one clear job

Implementation lead

Claude Code · Website repository. Built the site, pulled the work together, and asked Codex for help when needed.

Visual and review peer

Codex · Connected worktree. Made image assets, tried other UI ideas, reviewed the code, and checked the tests.

Keep every handoff visible

Show both agents the same site

Person running it. Both agents saw the task, repository, and live site. They reviewed the real result, not a written summary.

Split the work

Claude Code and Codex. Claude Code handled the main build and gave Codex smaller jobs where it had the better tools.

Review each other

Both agents. Each agent checked the other's design and code choices.

Check the checks

Both agents. Both agents tested the site, then reviewed each other's tests and security findings.

You still make the important calls

I could see every message, task, edit, and preview on one canvas.

Every task still needed a clear owner and proof that it was done.

I reviewed the final site and chose what shipped.

What changed

The site was better because the agents did more than split a checklist. They challenged work that was already done and caught different mistakes.

Testing and security improved the most. Each agent ran its own checks, then challenged the other agent's evidence before I accepted the result.

They also tried to hand boring work to each other. Because ownership was visible, those tasks did not quietly disappear.

Original post

All case studies · Download October