Auto-review Run Mode

Source: Cursor Date: 2026-05-29 URL: https://cursor.com/changelog/auto-review

Summary

Cursor’s auto-review run mode reduces agent interruptions by classifying tool calls (Shell, MCP, Fetch) into three buckets: immediately allowlisted, sandboxed, or routed to a classifier subagent that decides whether to proceed, suggest an alternative, or escalate to the user. Custom instructions can tune the classifier. The net effect is higher agent autonomy without removing the human-in-the-loop entirely.

Implications

• Direct signal for the agentic engineering patterns thread: classifier-gated autonomy is emerging as the practical middle ground between fully supervised and fully autonomous agent operation.
• The sandboxing path is relevant to security/Mythos: Cursor is operationalizing the idea that tool calls can be risk-tiered at runtime, which is a pattern that will propagate to other coding agents.
• The subagent classifier architecture is also a inter-agent trust data point—decisions about what an agent is allowed to do are increasingly delegated to another model rather than hard-coded policy.