San Francisco: Anthropic has added a new auto mode to Claude Code, its agentic coding tool, giving the AI the ability to decide which actions it can take independently, without waiting for developer approval.
The feature is currently in research preview. It is not a finished product but is available for testing.
Auto mode works by running an AI safety layer over every action before it executes. Safe actions proceed automatically. Actions flagged as risky or suspected of being triggered by prompt injection attacks, where malicious instructions are hidden inside content the AI is processing are blocked before they run.
The feature builds on Claude Code’s existing dangerously-skip-permissions command, which previously handed all decision-making to the AI with no safety filter. Auto mode adds that filter on top, making autonomous operation more practical for real development environments.
The shift is meaningful. Until now, developers using agentic coding tools faced a binary choice: approve every action manually or let the model run unchecked. Auto mode introduces a third option, supervised autonomy, where the AI itself determines what requires human sign-off.
Anthropic has not disclosed the specific criteria its safety layer uses to distinguish safe from risky actions. That detail will matter to enterprise developers before they adopt the feature at scale.
Auto mode follows two recent Claude Code launches. Anthropic shipped Claude Code Review in March, an automatic bug-catching tool for AI-generated code. It also launched Dispatch for Cowork, which lets users delegate tasks to AI agents working on their behalf.
Together, the three features outline Anthropic’s direction for Claude Code as less a coding assistant, more an autonomous developer that knows when to act and when to ask.
Previous
