Codex for Async Tasks: Offloading the Background Work

Six months into running this serverless stack, I have a clear pattern for delegating work. The interactive stuff stays in Claude Code. The long-running background jobs go to Codex. That split has saved me hours every week and kept my terminal free for the work that needs attention.

The Interactive vs Async Split

Claude Code is great when I need to review each change, negotiate the next step, or reason about a tricky bug. But some tasks do not need me in the loop. Bulk refactors across dozens of files. Content migrations. Test generation for an existing module. Those are async tasks, and Codex cloud agents run them in the background while I do something else.

Here is my current rule of thumb:

• Interactive (Claude Code in terminal): architecture decisions, debugging, writing new features
• Async (Codex cloud agents): bulk refactors, test generation, content migrations, dependency upgrades
• Never async: anything that touches production secrets or deploys to live environments

A Real Codex Task

Last week I needed to update about 40 blog posts to use a new category structure. The pattern was clear but the execution was tedious. I wrote a Codex task that looked like this:

Task: Migrate category structure on all posts under content/posts/.
Input: category mapping in content/categories.json
Output: rewrite each post frontmatter to new categories
Verify: no post loses a category, no new categories appear

I kicked off the task, went to a client call, and came back to a clean pull request. The branch had 40 file changes, a summary comment, and the verification logs. I reviewed, merged, done.

When Codex Is Worth It

Codex adds friction. You write a task spec, wait for the environment to spin up, then review the output. For a five-minute change, interactive is faster. The breakeven is around 30 minutes of expected work. Below that, stay in your terminal. Above that, describe the job and walk away.

The other big win is parallelism. I can have three Codex tasks running while I work on the hard problem. That is the skill soup in practice. Not one tool doing everything, but the right tool for each slice.

Read more about Codex in the OpenAI Codex documentation.