Anthropic Gives Claude Code an Auto Mode — With a Safety Leash Attached
Anthropic's new auto mode for Claude Code lets the AI execute development tasks with far fewer approval prompts. It ships alongside two other agent tools, but Anthropic hasn't disclosed the exact criteria its safety layer uses to greenlight actions.
Anthropic shipped auto mode for Claude Code on March 24, and it changes how the tool handles approval gates. Instead of pausing to ask the developer before every file edit, terminal command, or browser action, Claude Code now runs a continuous safety layer in the background and only interrupts when it judges an action too risky to proceed without human sign-off.
The pitch is speed. Developers using Claude Code have complained that constant confirmation prompts break their flow, especially on longer multi-step tasks. Auto mode is Anthropic’s answer: trust the model to keep moving, intervene only when something looks dangerous.
What counts as dangerous, exactly? That’s still murky. Anthropic has not published the specific criteria its safety layer uses to distinguish low-risk from high-risk actions, and TechCrunch noted the company hadn’t responded to requests for clarification by publication time. For teams considering wide deployment, that gap matters — you want to know whether auto mode would, say, push a commit without asking, or execute a database migration unprompted.
Auto mode currently works only with Claude Sonnet 4.6 and Opus 4. It rolls out to Enterprise and API users in the coming days.
The release isn’t standalone. Anthropic bundled it with two other launches:
Claude Code Review is an automated code reviewer that’s meant to catch bugs before they reach the codebase. It runs as part of the development loop rather than as a separate step, and the idea is that you get a second-pass review from the model without having to context-switch.
Dispatch for Cowork lets users hand off tasks to AI agents directly — delegating work rather than supervising every step. It’s a more explicit step toward the “AI employee” framing that Anthropic and its competitors have been pushing all year.
Together, these three features represent a deliberate move up the autonomy spectrum. Claude Code started as a terminal-native assistant that answered questions and generated snippets. It’s now executing multi-step workflows with reduced human checkpoints.
That trajectory is not unique to Anthropic. OpenAI’s Codex, Google’s Jules, and GitHub Copilot Workspace are all pushing in the same direction: fewer interruptions, longer autonomous runs, more action and less conversation. The competitive pressure is obvious.
The honest constraint is trust. Developers will adopt auto mode at the pace they trust the model — and trust requires knowing, concretely, what the model will and won’t do on its own. Anthropic’s refusal to document the safety criteria is a gap, not a feature. Publishing that logic, even at a high level, would accelerate adoption far more than any benchmark result.
For now, auto mode is a compelling upgrade for teams already deep in the Claude Code ecosystem. If you’re running repetitive scaffolding tasks or working inside a well-defined codebase, the reduction in prompts alone is worth testing. Just don’t deploy it to production pipelines until Anthropic clarifies what’s actually on the leash.