Claude Code or Codex: picking the agent per task

They are the same shape

Claude Code and Codex are both terminal agents: you give them a repository and a task, they read files, run commands and edit code, and they come back with a diff and a summary. Both are driven from a shell, both keep a session, both can be interrupted.

That similarity is why "which is better" is the wrong question. On a given day the difference that decides it is usually which one you have a subscription for, or which one has been less confused by this particular codebase.

Where they tend to differ in practice

Two things show up reliably, and both are about how you work rather than which model is smarter:

• Session style. One keeps a longer conversational thread and answers questions well mid-task; the other is happier being given a narrow, mechanical instruction and left alone.
• Review burden. The same task run twice produces two different diffs — different scope, different amount of incidental cleanup. Reading both is often faster than arguing about which agent is better.

Run both, then decide

The practical answer is to stop choosing. Give the same task to both agents in two worktrees — not the same checkout — and read the two diffs. It costs one extra session and it settles the question with evidence from your codebase rather than from somebody’s benchmark.

That is also the reason to run them in a workspace: two agents in one checkout will overwrite each other’s edits, and the failure is silent. One worktree each, and the comparison is free.