Thought Leadership
Anthropic is making the so-called Auto mode in Claude Code the default setting starting August 14. A classifier will now review every action instead of the user.
Anthropic had already introduced Auto mode as a test feature back in March, aiming to better balance speed and control during longer coding tasks. Instead of asking for confirmation on every single file or command action as before, a separate classification model now reviews each individual tool call before it is executed. If the classifier blocks an action, Claude Code first tries an alternative, safer approach or explicitly asks for clearer approval. After three consecutive or a total of 20 blocked actions within an interactive session, the system automatically switches back to manual confirmation mode; in non-interactive command-line use, the process instead aborts entirely in that case.
Anthropic justifies the move with its own research: in a controlled study involving 1,053 paid human test subjects, participants caught on average only 13.6 percent of intentionally embedded dangerous commands, a rate that fell to around 5 percent after roughly 50 consecutive confirmation requests. By comparison, the company’s own classifier blocked 937 of 1,053 crafted dangerous commands, or nearly 89 percent, but likewise let 116 commands through undetected. Anthropic explicitly stresses that the classifier does not eliminate the risk entirely either, and continues to recommend that users manually review larger changes to production systems.
Opt-in for enterprise customers, more pull requests among existing users
For Claude Enterprise customers, as well as for use via the Claude API, Claude Platform, Amazon Bedrock, and Google Cloud, Auto mode remains an opt-in feature for now; according to the company, a corresponding default rollout is planned there for the coming weeks. Administrators can also specifically set the default mode for their organization or disable Auto mode entirely, and individual users can switch between modes via a keyboard shortcut. Anthropic will no longer count the additional tokens needed by the classifier itself against existing usage quotas. According to the company, Team and Enterprise customers who already use Auto mode today create on average around 25 percent more pull requests than users in the previous manual mode.
(Editorial Team)
