Agents Need an Undo Button, Not More Approval Prompts A proposed safety framework for AI agents argues that reversibility, not approval prompts, should determine which agent actions require human oversight, citing the 1987 Sagas paper by Hector Garcia-Molina and Kenneth Salem as the design template. The framework sorts agent actions into four categories — reversible, compensable, delayable and irreversible — and notes that reversibility decays over time, citing Gmail's Undo Send as an example of a hold window that lasts up to 30 seconds before a message actually leaves. The argument holds that approval queues degrade into reflex clicks, so each tool should record how long its undo stays valid and what else can read its effect before then. Article contents The approval prompt problem