The rule. Give every action an AI takes on a user's behalf a working way back: undo for edits, version history for records, preview before send, and checkpoints with rollback for agents.
Why. People perform actions by mistake even when the software behaves perfectly; Nielsen's third usability heuristic requires "a clearly marked 'emergency exit' to leave the unwanted action without having to go through an extended process," with undo and redo as the canonical supports (Nielsen, 1994). AI systems add a second source of error, the system itself, so the guidelines for human-AI interaction (validated against 20 popular AI-infused products) devote a whole phase to being wrong, including guideline 8, "make it easy to dismiss or ignore undesired AI system services," and guideline 9, "make it easy to edit, refine, or recover when the AI system is wrong" (Amershi et al., 2019; Microsoft, n.d.). Agents raise the blast radius from one wrong suggestion to a batch of changed files, so current agent tooling ships state capture as core infrastructure: a checkpoint before each prompt is what lets users "pursue ambitious, wide-scale tasks knowing you can always return to a prior code state" (Anthropic, n.d.).
Seen in the wild. Claude Code checkpoints code state before each user prompt and offers /rewind to restore the code, the conversation, or both; the docs plainly warn that changes made by bash commands fall outside the net (Anthropic, n.d.).