Back to catalog
Control & Agency
Irreversible Actions
The AI takes a destructive action (sending email, deleting file) without confirmation.
3 principles3 checklist items2 simulation modes
What you will learn
Confirmation Modal
Undo Capability
Staged Execution
Run the simulation to see the failure in action.
Interactive Lab
Toggle between Bad and Good modes.
Anti-pattern mode
Simulation Status: Idlev2.4.0
Why it fails
AI agents that can interact with the world must not execute sensitive commands immediately. 'One click' efficiency is dangerous when the 'click' is a fuzzy natural language prompt.
Root cause
Optimizing for speed. Assuming the model 'understood' the intent perfectly. Lack of a 'Human-in-the-loop' confirmation step.
How to fix it
Confirmation Modal
Undo Capability
Staged Execution
Implementation checklist
- Require explicit confirmation for any external side-effect
- Show a preview of the payload (e.g., email draft) before sending
- Provide an 'Undo' window if possible