What the source did
- Gave both tools the same exception, code and logs for a legacy production bug; Claude Code traced the root cause.
- Had Codex run tests across modules while the fix wrapped up.
- Kept separate CLAUDE.md and AGENTS.md, plus a CHANGES.log recording status, files, remaining work and next step.
“Claude Code handles understanding. Codex handles execution.”— Sam Lai, DEV Community
Summary written by Agent Cookbook. See the original for the full context.
Copy the prompt
Ours Our version, based on what Sam Lai described. Not their wording. Fill in the [BRACKETS].
ROUTER (decide per task):
- Reasoning-heavy (root cause, business logic, change boundary) -> Claude Code
- Mechanical / parallel (tests, docs, repetitive edits, follow-ups) -> Codex
- Might switch mid-task -> write state to CHANGES.log first
CLAUDE CODE: Here are the exception, the logs and [FILES]. Trace the call chain and explain why it fails in this scenario. Propose the smallest fix and list follow-up work as bullet tasks in CHANGES.log.
CODEX: Read the latest CHANGES.log entry. Implement the listed follow-ups and add tests for [MODULES]. Run them and append Status, Files modified and Tests results.
Approval rule
🛡
You make the final call when the two agents disagree; neither pushes to shared branches.
Set it up
- Write the router rules at the top of CHANGES.log.
- Start reasoning tasks in Claude Code with tight context: the exception, key logs, the interface.
- Hand mechanical follow-ups to Codex via the log.
- When answers differ, compare reasoning yourself before choosing.
Common failures
⚠
- Dumping whole folders as context makes both tools generic. Keep inputs tight.
- Using the heaviest model for routine work slows everything down.
Variations
- Reverse roles for a week and compare.
- Add a third pass: one agent reviews the other's diff.
Get the 10 best agent setups. Free.
Plus new recipes every week. One email, no spam.
Unsubscribe anytime. Privacy