It's been wild to witness the emerging divergence in behavior between Grok / Claude.

Grok Bot is finding success leaning into default-permissive: it almost never asks for permission, hides tool calls, has full virtual computer use, very little scoped access per agent. It makes you feel like you can do anything.

Claude is trending the other way: increasingly restricted and unpredictably frustrating to use in the name of preventative safety, even when the threat is a false positive. This inevitably leads to a bad UX like the one below:

Anthropic have dug themselves into a bit of a hole; their positioning has forced them to take this approach, and leaning into the default-permissive model will be a change in their safety-first modus operandi.

I'd love to be a fly on the wall in the Anthropic office listening to the deliberation that must be happening around this: loosen up safety in the name of UX in order to compete?