Anthropic hands Claude Code more control, but keeps it on a leash

Source: TechCrunch

Image Credits: Jagmeet Singh / TechCrunch

The Challenge of “Vibe Coding”

For developers using AI, “vibe coding” currently means either babysitting every action the model takes or letting it run unchecked. The industry is moving toward tools that can act without waiting for human approval, but finding the right balance between speed and control remains a key challenge. Too many guardrails slow development; too few make systems risky and unpredictable.

Anthropic’s Auto Mode

Anthropic’s latest update introduces auto mode, now available in a research preview. The feature lets the AI decide which actions are safe to execute automatically, while still applying safety checks.

• Safety layer: Reviews each action before it runs, looking for risky behavior the user didn’t request and for signs of prompt injection (malicious instructions hidden in content).
• Execution: Safe actions proceed automatically; risky actions are blocked.
• Underlying tech: Extends Claude Code’s “dangerously‑skip‑permissions” command with an added safety layer.

How It Fits Into the Autonomous Coding Landscape

Auto mode builds on a wave of autonomous coding tools from companies such as GitHub and OpenAI that can execute tasks on a developer’s behalf. It pushes the decision of when to ask for user permission from the developer to the AI itself.

Anthropic has not disclosed the exact criteria the safety layer uses to differentiate safe from risky actions, a detail many developers will likely want before adopting the feature widely. (TechCrunch has reached out for more information.)

Related Anthropic Features

• Claude Code Review: An automatic code reviewer that catches bugs before they reach the codebase.
  Read more
• Dispatch for Cowork: Allows users to assign tasks to Claude agents from anywhere in the Cowork environment.
  Learn more

Availability and Recommendations

Auto mode will roll out to Enterprise and API users in the coming days. It currently works only with Claude Sonnet 4.6 and Opus 4.6. Anthropic recommends using the feature in isolated environments—sandboxed setups kept separate from production systems—to limit potential damage if something goes wrong.