Cursor SDK Adds Agent Steering, Background Subagent Results, MCP Hints, and Custom System Prompts

Key Info

Cursor SDK now supports steering running agents, background subagent result reporting, MCP annotations on custom tools, and custom system prompts, with support for local TypeScript and Python agents.

Highlights

• run.steer() injects your message into the next turn; a mid-task subagent moves to the background and keeps working
• Background subagents report results back to the parent as a follow-up turn on the same run, so stream() and wait() continue until all subagents finish
• Custom tools can carry MCP annotations like readOnlyHint and destructiveHint, letting the model distinguish a lookup from a delete before calling
• Cursor's system prompt can be replaced with your own while rules, skills, and tool schemas still load; rolling out account by account
• Steering and system prompts work with local TypeScript agents; background subagent results return locally in TypeScript and Python