AI agents promise hands-off automation, but most real-world teams are still wrestling with a basic question: how much autonomy do you actually want to give your code?