"Ask Every Time" is the opposite of safe

Ideas Feedback
1

My Cursor shows three possible settings for the Auto-Run Mode:

  • Ask Every Time
  • Use Allow List
  • Runs Everything

Supposedly “Ask Every Time” is safest. But if you study humans you know this is absolutely not true. The irritation of having to click “Run” 50 times causes an extremely predictable human effect…. people stop reading and just click “Run”. The system rapidly exceeds their patience and their cognitive abilityto keep up.

If you study the behavior of people who are asked to change their password often (a similar, cognitively demanding task), they begin choosing simpler and simpler passwords to lower the cognitive burden.

I would argue that the allowlist is safest because it reduces the cognitive load on people so they are more likely to actually look at concerning items. I beg you to rethink this. I’d simply turn it off, but “Ask Every Time” is the only setting that allows me to make updates elsewhere on my linux server.

1 Like
3

I realize I didn’t add a suggestion for an alternative and I want to be constructive.

Add a 4th mode called “Allowlist plus Analyze”. This mode would:

  • Combine the allowlist with a “Mini” model to inspect the action that is being requested for aspects like: security, undo-ability, etc.
  • This would be slower, but give a warning to indicate whether the user should take a closer look at the given propose agent action
  • Allow the user to click a button “Deep Analyze” to think more deeply, explain why the action is being proposed and what safety measures address the concern raised by the quick analyze.

The goal of this feature is to call the humans attention to exceptions rather than asking them to pay attention to literally every single agent action.