Cursor subagents

Subagents are isolated agent sessions the main Cursor agent can spawn via the Task tool to handle a complex, multi-step piece of work autonomously. Each subagent runs with its own context window and returns a single summary message when done — so broad exploration or parallel experiments don't fill up the parent conversation. Common types: explore (read-only codebase search), generalPurpose (multi-step research), shell (git / command execution), browser-use (web automation), cursor-guide (questions about Cursor itself), best-of-n-runner (isolated git worktrees for parallel attempts).

The sweet spot is delegation: "find everywhere we call the old API" (one explore subagent), "spin up 3 variations of this function and pick the best" (three best-of-n-runners). Not worth it for single-file edits — the overhead eats the benefit. Mental model: worker threads for the agent.