Cursor
@cursor_ai
Auto-review mode is now available in Cursor. It allows agents to run tool calls with fewer approval prompts and safer execution. https://t.co/GZQX89mgmq
96retweets1.7klikes
View on XCursor Auto-review Mode Uses Subagents to Cut Approval Fatigue
Cursor, an AI-native code editor, introduced Auto-review to streamline how its agents handle tool calls. The mode categorizes actions into three tiers: allowlisted commands run immediately, sandboxed tasks execute within secure boundaries, and all other requests route to a secondary classifier subagent for autonomous evaluation.
This update addresses the "approval fatigue" that often stalls complex development workflows. By using a dedicated reasoning model to judge the safety of Shell and MCP tool calls, Cursor reduces the need for manual human clicks. This shift parallels OpenAI's Auto-review for Codex, signaling an industry-wide push toward unattended agent operations.
You can enable the new mode under the agent settings menu and provide custom instructions to steer the classifier's decision-making logic. The feature is available now for all Cursor users, applying to Shell, MCP, and Fetch tool calls. This builds on Cursor's cross-platform agent sandboxing to provide a more autonomous development experience.
Still wondering? A few quick answers below.
Auto-review is a new run mode for the Cursor AI code editor that allows agents to work for longer periods with fewer manual approval prompts. It uses a tiered system to handle tool calls, delegating complex decisions to a specialized classifier subagent to maintain safety while reducing the need for human intervention.
When an agent attempts a tool call that is not on an allowlist or cannot be sandboxed, Cursor routes the action to a classifier subagent. This secondary agent analyzes the request to decide whether to allow the call, try a different approach, or ask the user for manual approval.
Auto-review mode currently applies to Shell, MCP, and Fetch tool calls. Shell commands involve terminal interactions, while MCP refers to the Model Context Protocol used for connecting to external data and tools. Fetch calls allow the agent to retrieve information from web sources or remote APIs during the coding process.
Yes, users can steer the classifier subagent by providing it with custom instructions. These instructions help define the agent's decision-making logic for specific workflows. You can configure these settings and toggle the run mode by navigating to the Agents section within the Cursor application settings menu.
Auto-review mode is available as a new run mode within the Cursor editor as of May 2026. Any user with access to Cursor's agent features can enable this mode in their settings to reduce approval fatigue during multi-step coding tasks involving terminal commands, external tools, or web-based data retrieval.
