Claude Code users approve 93% of permission prompts. That is a control-design problem.

What Anthropic reported

Anthropic says Claude Code users approve 93% of permission prompts. The company describes the resulting approval-fatigue problem and the tradeoff between sandboxing, manual approvals and the unsafe --dangerously-skip-permissions option.

Its response was auto mode: classifiers decide whether some requested actions can proceed without a manual permission prompt. Anthropic is explicit that the approach improves the tradeoff but does not eliminate risk.

Why this matters beyond one coding tool

A permission popup is only useful when it appears at the right boundary and gives the reviewer enough context to make a real decision. If every read, write and shell command looks equally urgent, approval becomes muscle memory.

The better policy question is narrower: who is acting, on whose authority, against which resource, with which arguments, in which environment, and how consequential is the action?

Where RaksHex fits

RaksHex's relevant design principle is semantic, pre-execution authorization. Routine low-risk actions can be allowed by policy, obviously prohibited actions can be denied, and consequential actions can require approval with the action, arguments, identity and delegated authority attached.

That does not mean RaksHex replaces Claude Code's own sandboxing or Anthropic's safety classifiers. It sits at a different boundary: deciding whether a consequential action routed through RaksHex should execute.

Source

Anthropic Engineering, March 25, 2026. The 93% approval figure and discussion of sandboxing, permission bypass and auto mode come from Anthropic's article.