Reversible vs irreversible agent actions: should humans approve every risky step?

rcleffke · reddit · 2026-11-18

A Reddit discussion proposes classifying agent actions as reversible (drafting, searching, summarizing) vs irreversible (sending, deleting, paying, publishing, changing records), and requiring human approval enforced in code for the latter. Open questions include approval fatigue and case-by-case vs explicit lists. The poster runs a small agent risk-check service.

Related event: Agents need human approval for irreversible actions(2 posts)→

Original post →

More from coding & agent

coding & agent channel →