Permission level

A permission level is how much the agent may do without asking: plan-only, ask-per-action with an allowlist, auto-accept edits, or no prompts at all. It is a blast-radius control, not a verification control — raising it does not make the work more correct, only less interrupted. Anthropic's own documentation says the unrestricted mode offers no protection against prompt injection, which is the honest framing of what is being traded. This is the setting people change most often and reason about least. Pair every step up with a gate that does not depend on the model's own account of itself.