Anthropic is rolling out a new “auto mode” for Claude Code that takes over most of the annoying permission prompts developers usually have to click through, and it’s now live for Enterprise customers and available via the API.

Traditionally, Claude Code stops and asks for approval before it runs shell commands or edits files, which is safe but quickly turns into click-fatigue when you’re iterating on a feature or refactoring a large codebase. Anthropic’s own data shows developers were approving roughly 93% of these prompts anyway, often without really reading them, which basically turns a safety feature into a reflex. Auto mode is designed as a middle ground between “approve everything manually” and “turn all permissions off and hope for the best.”

With auto mode on, Claude uses a separate classifier to review each action before it actually runs. Routine, low‑risk stuff like tweaking local files or running harmless commands just goes through automatically, so your workflow isn’t constantly interrupted. When something looks risky—think mass file deletions, production deployments, credential access, or anything that smells like data exfiltration—the classifier blocks it and nudges Claude to try a safer approach. If Claude repeatedly keeps pushing towards dangerous actions, auto mode will eventually fall back to asking you for explicit permission again, so there’s still a human in the loop for genuinely sensitive decisions.

From a developer’s point of view, the experience should feel closer to working with a focused pair‑programmer than a needy intern constantly tapping you on the shoulder. You can kick off longer tasks—like large refactors, codebase cleanups, or multi‑step debugging sessions—and let Claude keep going without babysitting every single shell call. The command to start using it from the CLI is simple: update to the latest version and run claude --enable-auto-mode. For teams that aren’t ready to trust it yet, admins can still turn auto mode off at the org level via managed settings or OS policies.

Anthropic is pretty explicit that this is about reducing risk, not eliminating it. The classifier can sometimes misjudge edge cases—letting a risky action slip through or blocking something harmless—so the company still recommends running Claude Code in sandboxed or isolated environments, not directly on machines with production credentials. For Enterprise and API users, though, auto mode is a clear signal of where AI coding tools are headed: agents that can make more of their own decisions by default, while safety systems quietly watch in the background.