Auto mode made code review the last human gate

Anthropic made auto mode the default in Claude Code this month, and I think that is the right call. But it moves the weight of human judgment somewhere specific: code review is now the last human gate, and it matters more than it did last week.

The number nobody is talking about

In Anthropic's own study, human reviewers caught 13.6% of dangerous commands. The classifier caught 89%.

Sit with that for a second. We spent two years telling ourselves the human approval prompt was the safety layer. Turns out most of us were just pressing enter.

Why the change is still right

The approval prompt trained people to stop reading. When every tool call asks for permission, permission stops meaning anything. A classifier that actually reads every tool call beats a tired engineer clicking approve for the 40th time before lunch.

The part I would put on a wall

Anthropic said something else in the same announcement: classifiers cannot eliminate risk. Keep human review for production changes.

So the approval prompt is dead. Code review is not. If anything, review just became the only human gate left between an agent's output and production.

Did your review process keep up?

Teams on auto mode ship around 25% more PRs. The question worth asking on your team is whether your review process got 25% better to match. More diffs, written faster, with less human attention at the moment of writing, means the review is doing more of the work than it used to.

Takeaways

• Human approval prompts were a weak safety layer; people stopped reading them.
• A classifier reading every tool call is a better default than approval fatigue.
• Classifiers do not eliminate risk. Production changes still need human review.
• If your PR volume went up, your review process has to scale with it.