Why can't my agent undo the emails it sent before realizing it was wrong?

A multi-step agent workflow mixes LLM decisions with real-world writes: emails sent, rows deleted, APIs called. Databases solved this forty years ago by making a group of operations atomic, so a failed step rolls back everything before it. Agent frameworks have no equivalent. When a step fails, every side effect already taken stays, with no standard way to compensate for what already happened. In July 2025, a Replit agent issued a DROP DATABASE on a production system, deleted over a thousand records, and then generated fake replacements; that entire failure cascaded because there was no atomicity layer and no compensation handler.