Use the cursor-agents MCP tools to break work into sub-tasks and delegate them to Cursor agents.
Synchronous delegation (need result now)
Use cursor_run when you need the output before continuing:
cursor_run(
prompt="<full task description with all context the agent needs>",
cwd="/path/to/project" # optional, defaults to current directory
)Returns { text: "<agent output>", stop_reason: "..." }.
Background delegation (parallel work)
Use cursor_task_submit to fire off tasks and continue working:
- Submit one or more tasks:
`` cursor_task_submit(name="Refactor auth module", prompt="...", cwd="...") cursor_task_submit(name="Write tests for API", prompt="...", cwd="...") ``
- Check status:
`` cursor_task_status(task_id="<id>") ``
- Fetch result when completed:
`` cursor_task_result(task_id="<id>") ``
- See all tasks:
`` cursor_tasks_list() ``
Writing good prompts for Cursor agents
- Include full context — the agent has no knowledge of what you've done so far
- Specify the working directory if the task involves files
- Be explicit about what "done" looks like (files to edit, tests to pass, output expected)
- If the task spans multiple files, list the key ones