Cursor Adds Auto-review Mode to Streamline Agent Workflows

TL;DR

• Cursor has launched a new Auto-review run mode designed to enhance AI agent workflows.
• This mode allows agents to operate for extended periods with reduced user intervention and increased safety.
• Shell, MCP, and Fetch tool calls are specifically supported by the Auto-review functionality.
• Users can configure and customize this feature through Cursor's settings and custom instructions for the classifier agent.

Cursor is leveling up its AI agent capabilities with the introduction of a new Auto-review run mode. This innovative feature is designed to significantly streamline workflows by allowing AI agents to operate for longer durations without constant user prompts, while simultaneously enhancing the safety of their tool interactions. This means developers can expect a more fluid and efficient AI-assisted coding experience.

The Auto-review mode directly impacts how agents handle various tool calls, specifically Shell, MCP, and Fetch tool calls. For allowlisted commands, execution is immediate, ensuring critical operations proceed without delay. Similarly, calls that are suitable for sandboxing are automatically run within a secure sandbox environment, providing a layer of protection against unintended consequences.

For all other agent actions, a sophisticated classifier subagent takes the reins. This intelligent component analyzes each action and makes a crucial decision: whether to approve the call outright, attempt a different approach if the initial one seems problematic, or, in situations requiring human judgment, request explicit user approval. This intelligent decision-making process aims to balance autonomy with user control.

Users have direct control over how this new feature behaves. The Auto-review run mode can be easily configured within Cursor's settings, accessible via Settings > Cursor Settings > Agents > Run Mode. Furthermore, for those seeking even finer-grained control, the classifier agent's behavior can be further customized through the use of custom instructions, allowing developers to tailor the AI's decision-making process to their specific project needs and preferences.

This enhancement is particularly beneficial for complex or lengthy tasks where an agent might need to execute a series of commands or make multiple tool calls. By reducing the need for manual approval at each step, Auto-review allows the agent to maintain momentum, leading to faster task completion and a less interrupted development cycle. The added safety measures ensure that these extended operations remain secure.

To learn more about how this feature works and how to implement it effectively within your development environment, you can consult the detailed documentation. The ability to customize the classifier agent is a powerful tool for developers who want to ensure their AI agents operate exactly as intended, providing a tailored and efficient coding assistant. Read more about Terminal Run Mode here.

Summary

• Cursor's new Auto-review run mode boosts AI agent efficiency by reducing approval prompts.
• It enhances safety for Shell, MCP, and Fetch tool calls through immediate execution or sandboxing.
• A classifier subagent intelligently manages other actions, deciding whether to approve, retry, or seek user input.
• Users can configure the run mode and customize the classifier agent's behavior via settings and custom instructions.

Source: Auto-review Run Mode · Cursor