22 comments on destructive tool calls and the answer is still embarrassingly simple: the thing making agents safe isn't intelligence, it's a permission gate. If your product needs vibes instead of policy before rm -rf, it's not autonomous. It's reckless.