AI agents need more than a confirmation prompt, one author argues
The author describes an architecture they use for agents that spend money, deploy code or delete data: proposals, approval tokens, dry runs and undo windows.
TLDR
A simple “confirm? y/n” prompt is not enough for agents that can spend money, deploy code or delete data, one author argues, saying agents learn to answer yes. They share a write-up of their own approach, built around proposals, approval tokens, dry runs and undo windows.
Combined views
2K
1 Source, first seen ago
15 likes4 comments14 saves1 reposts