Fewer clicks, more control? Claude Code in Auto Mode by default

Anthropic announced it will enable Auto Mode by default in its Claude Code AI development tool starting Aug. 14 for Pro, Max and Team license holders, the company said. The change allows the system to execute shell, git and tool calls independently without requesting user approval at every step, relying on a classifier to evaluate the security of each action. The classifier blocks steps it deems unsafe, such as potentially irreversible or destructive changes, and triggers a switch to manual approval after three consecutive blockages or 20 blockages in a session. Anthropic said a controlled study showed humans recognized only 13.6 percent of intentionally injected dangerous commands, while Auto Mode blocked 89 percent. The company also cited user permission fatigue, noting that 97 percent of AI requests were approved reflexively. Security mechanisms include hard denies for actions like potential data exfiltration and prompt injection screening. Anthropic said users of Auto Mode are expected to deliver about 25 percent more pull requests, referencing use at Adobe, Nuro, Gusto and Garner Health. The classifier's additional token costs will no longer be charged for the affected plans. Enterprise, API and cloud platform users will be switched to the default in the coming month.