Claude Code Introduces Auto Mode for Safer, Longer-Running Tasks

Today, Claude Code is rolling out a new feature called "Auto Mode," designed to provide a safer and more efficient way to run longer tasks without the frequent interruptions of manual permission checks. This mode, available as a research preview for Team plan users, introduces automated decision-making with safeguards to monitor actions before they execute. It's set to roll out to Enterprise and API users in the coming days.

Why Auto Mode Matters

Claude Code’s default permissions are designed to be conservative: every file write or bash command requires explicit approval. This approach ensures safety but can be cumbersome for long-running tasks that require frequent human intervention. Some developers opt to bypass permission checks using --dangerously-skip-permissions, which can lead to dangerous and destructive outcomes, especially outside isolated environments.

Auto Mode offers a middle ground by allowing longer tasks to run with fewer interruptions while introducing less risk than completely skipping permissions. Before each tool call runs, a classifier reviews the action for potential risks such as mass deleting files, sensitive data exfiltration, or malicious code execution. Safe actions proceed automatically, while risky ones are blocked and redirected.

How Auto Mode Works

• Classifier Review: Each tool call is reviewed by a classifier to check for potentially destructive actions.
  • Safe Actions: Proceed without user intervention.
  • Risky Actions: Blocked and redirected, with Claude taking an alternative approach if possible.
• User Prompts: If Claude insists on risky actions that are continually blocked, it will eventually trigger a permission prompt to the user.
• Risk Management: While Auto Mode reduces risk compared to --dangerously-skip-permissions, it doesn't eliminate it entirely. The classifier may still allow some risky actions or block benign ones due to ambiguous user intent or lack of context.

Implementation Details

Auto Mode is available in Claude Code as a research preview for Team plan users today and will soon roll out to Enterprise and API users. It works with both Claude Sonnet 4.6 and Opus 4.6. Here are some key implementation details:

• Token Consumption: Auto Mode may have a small impact on token consumption, cost, and latency for tool calls.
• Managed Settings: For admins, you can disable Auto Mode in the CLI and VS Code extension by setting "disableAutoMode": "disable" in your managed settings. It is disabled by default on the Claude desktop app.

Under the Hood

The core of Auto Mode lies in its classifier, which uses machine learning to evaluate each action before it runs. The classifier is trained on a dataset of actions labeled as safe or risky based on their potential impact. Here’s a breakdown of how it operates:

• Data Collection: The classifier is continually updated with new data to improve its accuracy.
• Action Evaluation: Each tool call is evaluated for specific risk factors such as file operations, network requests, and code execution.
• Decision Making: Safe actions are approved, while risky ones are blocked or redirected. If Claude persists in requesting risky actions, it will eventually prompt the user for manual approval.

What to Watch

While Auto Mode represents a significant step forward in balancing safety and efficiency, there are a few things to keep an eye on:

• Accuracy Improvements: The classifier’s accuracy will continue to improve over time as it receives more data.
• User Feedback: User feedback is crucial for identifying edge cases and improving the overall experience.
• Security Best Practices: Even with Auto Mode, it’s essential to follow best practices for secure development and use isolated environments when necessary.

Auto Mode in Claude Code offers a pragmatic solution for developers looking to run longer tasks with fewer interruptions while maintaining a high level of safety. As this feature rolls out more widely, we can expect to see significant improvements in developer productivity and security.