Which of your agent's actions can't be undone, and does a person approve each one?
retained text
Here's a scenario: an agent with access to email and a calendar gets a vague request, decides it knows what you meant, and sends a message to a customer. A draft would have been fine, since you can edit it or throw it away. A sent email is out there for good. The distinction I keep coming back to is reversible versus irreversible actions. Drafting, searching, summarizing and proposing changes are all cheap to get wrong. Sending, deleting, paying, publishing and changing a customer record are not. Many setups give the agent both kinds of permission in one lump. The practice I've been thinking about: list every action your agent can take, mark the ones that can't be undone, and require a person to approve each of those before it runs. Curious how others draw the line: - Do you have an explicit list of irreversible actions, or is it decided case by case? - Is approval enforced in code (the tool simply can't run without it), or just in the prompt? - Which action surprised you by being harder to undo than you expected? - Do approvals get annoying enough that people start clicking through them? Disclosure: I run a small AI agent risk-check service, which is why this is on my mind. No link in the post; I'll put it in a comment per the sub rules.
document fetched 2026-11-18 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1xyol15/which_of_your_agents_actions_cant_be_undone_and