Claude Code Auto Mode is now the default in Claude. It’s also better than human performance.

From August 14th, Anthropic will be making Claude Code’s “auto” mode the default for Pro, Max, and Team users. The agent operates autonomously, completing tasks without needing approval for each step (except when performing an irreversible or destructive action). In a test involving 1,053 participants, we found that auto mode stopped 89% of unsafe actions from occurring, versus human review which allowed 97% of requests for approval to pass without inspection, and only identified 13.6% of true dangers.

Human oversight has been formally shown to be less safe than the model’s internal protections. This is a sea change moment for our industry, as we shift toward safety that relies on the model’s internal protections instead of approval buttons. Most future generations of agentic products will be built this way.