Click, Click, Oops

Your coding agent prompts for approval on every action, including file reads. Users approve reflexively within a second, and one of them approved a recursive delete without registering it. What is the right redesign?

Remove interactive prompts and rely on a complete audit log plus the ability to revert from version control after the fact

Tier prompts by consequence: auto-allow reversible read-only actions, and prompt only on state-changing ones, showing the exact command and arguments

Ask once at session start for a blanket approval covering the whole task, so the user makes one considered decision instead of forty rushed ones in a row

Have the model rate each action's risk and raise a prompt only for the ones it labels dangerous, keeping the human out of the routine calls entirely