Claude Code Auto Mode is Anthropic's latest attempt to let developers move faster without fully taking the brakes off safety.

The feature introduces a new permissions mode in Claude Code where the assistant can approve or reject many actions on the user's behalf.

Instead of prompting for every file write or shell command, Auto Mode runs an internal classifier before each tool call, scanning for destructive behaviour such as mass file deletions, sensitive data exfiltration or execution of malicious code. Actions judged safe go through automatically, while anything flagged as risky is blocked and Claude is nudged to try a different approach or escalate to a user approval prompt.

Anthropic has launched Claude Code Auto Mode as a research preview for Team plan users, with rollout to Enterprise and API customers expected in the coming days. The system currently runs its classifier on newer Claude models such as Sonnet 4.6 and Opus 4.6 and is not enabled for older versions or most third-party integrations. Developers can toggle Auto Mode via the Claude desktop app, command line tools or supported code editors, while administrators retain the option to disable it centrally through organisational settings.

Anthropic pitches Claude Code Auto Mode as a "middle path" between manually approving every change and using the far more dangerous --dangerously-skip-permissions flag. Longer refactors, builds and multi-step tasks can now run with fewer interruptions, though Anthropic acknowledges a modest increase in token usage, cost and latency due to the extra safety checks. The company stresses that the classifier will not catch every risky action and may at times block legitimate ones, so it recommends using Claude Code Auto Mode only in controlled or isolated environments, particularly away from production systems.

Claude Code Auto Mode marks a meaningful step toward more autonomous coding assistants, offering developers a smoother workflow without completely sidelining human oversight. Yet Anthropic's own warnings about misclassifications and the need for sandboxed deployment underline that this is still an experiment rather than a finished safety system. For now, Claude Code Auto Mode looks less like a licence to let AI roam free and more like a carefully leashed companion that development teams must still monitor closely.