3.3 Cloud Tasks: Fire-and-Forget PRs

Not everything needs you sitting at a terminal watching it happen, and treating every task that way is its own kind of inefficiency. Cloud tasks hand a piece of work to Codex running in a hosted container instead of your local machine — you describe the task, close your laptop if you want to, and come back to a finished pull request instead of a stalled session waiting on you.

This is genuinely different from a background process on your own machine. Codex sets up the environment itself, detecting your dependencies and build steps automatically, and container caching cut typical task completion time dramatically compared to earlier versions — OpenAI's own numbers put it around a 90% reduction from spinning up a fresh environment every time. The result lands as an actual pull request with a description of what changed and why, ready for human review, not a raw diff dumped in a chat window you now have to translate into a PR yourself.

The judgment call is knowing what belongs in the cloud versus what you should do locally and interactively. A well-scoped, clearly-describable task — "update this dependency across the codebase and fix anything that breaks," "add tests for this module," "implement this issue exactly as described" — is a strong cloud-task candidate, because you can describe the goal precisely enough that you don't need to watch it happen step by step. Something exploratory, where you genuinely don't know the right approach yet and want to steer as you go, belongs in an interactive session instead — plan mode locally, not a fire-and-forget cloud task.

Work moves between surfaces without losing context, too — start a task in the cloud, pull it into your IDE to keep iterating once the PR lands, or vice versa. That continuity is arguably the bigger shift here: not "the cloud is faster" so much as "local and cloud are the same ongoing thread of work now, not two disconnected tools."