Skip to content

Pick your style: Denkbaar

Book a conversation

Articles

Knowledge

Claude and Codex together: Claude directs, Codex builds, Claude reviews

Two coding agents from two makers, each with its own role. What it brings, and where it went wrong.

Two developers at separate laptops pass a task card and a result between them.
Give building and reviewing to two different agents. Illustration: Denkbaar.

We work with two coding agents: Claude Code from Anthropic and Codex from OpenAI. Claude directs and checks, Codex programs. Sometimes that is also cheaper.

Plain technical English Claude Code splits the work and writes a task for each part. Codex CLI carries out the task in its own git worktree. Claude reads the result and approves it or sends it back.

Why it matters

An agent that reviews its own work does not see its own mistakes. A second model, from another maker, looks differently. And each model has its own strengths: Claude cannot draw images itself, but it can direct Codex or an image model.

A blue task folder connects to a separate work folder beside a second laptop.
Claude writes the task and Codex builds in its own work folder. Illustration: Denkbaar.

How it works

backgood

Claude: plan and task

Codex: builds in its own worktree

Tests and check

Claude: review

Merge

Claude writes the task, Codex builds and tests, Claude rejects or approves.

This is how it grew at Denkbaar, according to our own commits and sessions:

  • March 2026: a first skill in which Claude calls Codex CLI as a second reviewer.
  • May 2026: Codex becomes our second coding agent; at the end of May, the first commit that Codex co-wrote.
  • October 2026: we make designs in Claude Design; Codex rebuilds them. Codex cannot read those designs itself, so a Claude agent first puts the screens into the project as files.
A reviewer studies two columns of code blocks on a screen.
Let the reviewer read the code itself. Illustration: Denkbaar.

What you do

  1. Give each agent one role. Building and reviewing go to two different agents.
  2. Write the task so that it stands on its own. Codex cannot ask back.
  3. Arrange the report back. In October 2026 we counted 60 Codex starts in our sessions. After a finished run, it was often silent for 25 to 80 minutes, because nobody picked up the result. The problem was not the building, but the report back.
  4. Let the reviewer read the code, not the builder’s explanation.

What we can do with you: we set up a team of agents that builds and reviews, with clear roles and reports. Book a conversation.

Sources

  1. openai/codex (Codex CLI, Apache-2.0)
  2. Anthropic: Building effective agents

Topics: agentscodingclaudecodex