AFK

AFK is the pattern where you hand the agent a task and walk away. No approvals, no mid-run steering; the agent runs to completion on its own and you read the result later. It needs a permissive permission mode so nothing blocks waiting for a click, and usually a sandbox or a worktree so a bad run can't damage anything that matters.

This is where the leverage of AI coding lives. One person can have several AFK runs going while doing something else, and the runs happen overnight or during meetings. The failure to watch for is quiet: the agent hits an ambiguity, picks an answer, and spends the next hour building tidy, internally consistent work on top of it. Nothing crashes. You only find out when you read the diff.

Make the run safe to leave. Do the thinking before you go: a grilling pass and a written spec take the guesses off the table. Put automated checks in the loop so failing tests stop the run instead of you. And make the finish line a branch or a pull request you can read through, never a merge that already happened. AFK doesn't make human review optional; it batches all of it into one sitting at the end, so what comes out of the run has to be worth reading.