Agent actions that aren't on your allowlist or can't be sandboxed go to a classifier subagent. This separate agent decides whether to allow the tool call, try a different approach, or ask you for approval. Learn more:

→ View original post on X — @cursor_ai,