Anthropic defaults Claude Code to autonomous tool execution with a classifier that blocks irreversible actions instead of prompting, eliminating approval friction while matching human safety decisions.
Summary
Removes manual approval overhead that causes reflexive clicking (97% approval rate) and cuts session interruptions, letting long-running tasks complete unattended. Pro/Max/Team users get this automatically starting August 14; Enterprise admins retain control.
Why it matters
Removes manual approval overhead that causes reflexive clicking (97% approval rate) and cuts session interruptions, letting long-running tasks complete unattended. Pro/Max/Team users get this automatically starting August 14; Enterprise admins retain control.
Implementation verdict
Replaces manual permission prompts with a learned classifier that routes tool calls through safety gates. Requires nothing—opt-in status remains on API/Enterprise/cloud platforms until next month. Safe to adopt now if you're on Pro/Max/Team; production data shows it catches 89% of dangerous commands vs. 13.6% human detection rate.
Sources
Dev Signal
Get briefs like this in your inbox — free, every weekday.
100+ sources compressed into one 4-minute read. Ranked, cited, implementation-ready.