For machine learning engineers, autonomous coding often feels like a binary choice: micromanage every step or hope the model doesn't break production. Anthropic is attempting to bridge that gap with a new "Auto Mode" for Claude Code, currently available in research preview.

Instead of waiting for human approval on every command, the system now evaluates its own actions. Before executing a task, internal safeguards scan for prompt injection attempts or unintended behavior. If the action passes safety checks, it proceeds automatically. If it looks risky, the system blocks it. This approach evolves beyond the existing "dangerously-skip-permissions" flag by adding an intelligent safety layer rather than removing guardrails entirely.

The feature aligns with broader industry moves toward agentic workflows, similar to recent offerings from GitHub and OpenAI. However, Anthropic shifts the permission decision from the user to the model itself. While the specific criteria for distinguishing safe versus risky operations remain undisclosed, the update follows recent launches like Claude Code Review and Dispatch for Cowork.

Auto mode is rolling out to Enterprise and API users shortly. It requires Claude Sonnet 4.6 or Opus 4.6. Anthropic advises deploying this capability only within isolated, sandboxed environments to mitigate potential damage if the safety layer fails. Until the company clarifies how the safety model makes decisions, widespread adoption may require cautious testing. For now, it represents a significant step toward reducing friction in AI-assisted development without sacrificing control.

Source: TechCrunch