Anthropic is turning Claude Code’s auto mode on by default
Anthropic will make "auto mode" the default setting in Claude Code, its AI coding tool, for Pro, Max and Team account holders from 14 August. The change means the tool will act with less human oversight, proceeding through most actions automatically rather than pausing for approval at each step, and will only stop to ask permission when an action is judged irreversible, destructive, or reaching outside the user's own environment.
Anthropic first trialled the feature in March and says testing has shown it performs better than manual review: in a study of 1,053 paid testers, auto mode caught 89% of harmful actions compared with just 13.6% caught by humans, partly because users tend to approve prompts out of habit, accepting 97% of them. Claude Code head Boris Cherny said his team has used auto mode exclusively for months and would not want to return to manual prompts, while Anthropic said it has also added safeguards such as prompt injection screening and customisable "hard deny" rules to guard against risks like data exfiltration.
- Claude Code's auto mode becomes default from 14 August 2026
- Auto mode caught 89% of harmful actions versus 13.6% by humans
- New safeguards include prompt injection screening and hard deny rules
Both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
Supporters argue that requiring less manual approval for routine coding actions removes friction that slows developers down without adding much real safety, since most edits are low-risk and easily reviewed after the fact via diffs and version control. They see it as a natural maturation of the tool: as the system's judgement and guardrails improve, defaulting to greater autonomy lets engineers focus on architecture and review rather than repeatedly rubber-stamping small steps, while the option to tighten oversight remains available to those who want it.
The case against
Sceptics contend that defaults matter precisely because most users never change them, so shifting the baseline towards autonomy exposes a much larger population of developers to the consequences of unreviewed actions, including subtle bugs, security flaws or unintended file changes that surface only after damage is done. They value a human checkpoint before, not just after, consequential actions are taken, and worry that normalising reduced oversight could erode developers' understanding of their own codebases and make accountability harder to establish when something goes wrong.