Your AI Just Learned to Approve Its Own Actions

Anthropic shipped auto mode for Claude Code this week. Instead of asking permission for every file write and bash command, Claude now uses internal safety classifiers to approve or block actions on its own. It’s positioned as the middle ground between the default mode (where you approve everything) and --dangerously-skip-permissions (where you approve nothing and hope for the best).

I use Claude Code with custom permission rules — hooks that gate specific actions, allowlists for safe operations, blocks on anything destructive. Auto mode is essentially Anthropic building that logic into the model itself, so developers who don’t configure their own guardrails still get some.

The interesting tension here isn’t about this specific feature. It’s about where we’re heading on the autonomy spectrum. Every developer tool that uses AI agents is going to hit this exact design problem: too many permission prompts and nobody uses it, too few and someone loses a directory.

The honest answer is that there’s no universal setting. What’s safe depends entirely on what you’re working on, what you can afford to lose, and how much you trust the model’s judgment in your specific context. Auto mode is a reasonable default for people who haven’t thought about it yet. But if you’re doing serious work, you’ll still want your own rules.

The fact that enough developers were using --dangerously-skip-permissions to make Anthropic build a safer alternative tells you everything about how the autonomy conversation is actually going.

Source: The Verge