Cursor Launches Auto-Review Mode for Autonomous Tool Execution
Key Takeaways
- ▸Auto-review mode reduces friction by executing pre-approved tool calls automatically while maintaining security through sandboxing
- ▸A classifier subagent intelligently decides which calls to allow, deny, or escalate to the user based on risk assessment
- ▸Users retain full control through customizable settings and can provide instructions to steer the classifier's behavior
Summary
Cursor has introduced Auto-review mode, a new run mode that enables the AI-powered code editor to execute tool calls with greater autonomy while maintaining safety guardrails. The feature applies to Shell, MCP (Model Context Protocol), and Fetch tool calls—allowlisted calls execute immediately, sandboxed calls run in an isolated environment, and all other actions are routed to a classifier subagent that determines whether to approve, suggest an alternative approach, or request user confirmation.
Auto-review reduces approval friction and streamlines workflows for developers who want Cursor to work more independently. Users configure the feature through Settings > Cursor Settings > Agents > Run Mode and can provide custom instructions to guide the classifier subagent's decision-making.
This release reflects the industry trend of AI agents becoming increasingly autonomous in handling routine tasks while maintaining robust safety and user control as core design principles.
