Delegate a task to a Cursor agent (Grok by default) running detached in the background. Returns immediately with an agent id — it does NOT wait. The agent shares none of your context, so prompt must be a self-contained brief: name the files, the goal, and how to verify. Project AGENTS.md, .cursor/rules, .agents/skills and .cursor/mcp.json are honored automatically. After spawning, run the returned attachCommand via Bash with run_in_background:true, passing the returned attachDescription VERBATIM as the Bash tool's description. The task panel labels the row from that description, so a generic one hides what the agent is doing from the user. The task's exit notifies you when the agent finishes — do not poll.

Send another turn to an existing agent, keeping its full conversation. Use this instead of spawning a fresh agent whenever the work builds on what that agent already did — it keeps the context and costs far less. The agent must have finished; use steer for a live run.

List agents with their state, elapsed time, cost and latest activity. One line each.

Show what one agent has done, as a compact digest — one line per tool call, not a transcript. Pass the since value from a previous call to get only new activity. The reply names the full transcript file; read it with shell tools only if the digest is genuinely insufficient.

Block until the named agents finish. This blocks YOUR whole turn, so prefer the background-task bridge from spawn — use this only when there is genuinely nothing else to do. On timeout it returns progress rather than nothing.

Inject a message into an agent's turn while it is still running, without killing its work. Use this to correct course mid-run — it is strictly better than stopping and re-prompting.

Cancel an agent's current run. The agent stays resumable via follow_up.