An orchestration design has a coordinator launch team-lead subagents, and each team lead launch worker subagents of its own. What should be expected when it runs in Cursor?

The principle — Cursor lets subagents launch child subagents, but the tree has a fixed depth. The main agent may launch subagents, and those direct subagents may launch subagents of their own; a subagent that was itself launched by a subagent cannot launch any further ones. Two levels of delegation, and then it stops.

Why the key is correct — Counting from the coordinator, the team leads are direct subagents of the main agent, so they spawn normally. The workers sit one level below that, which is precisely where launching is refused. The leads still run and still report back, so the failure is quiet: the leaf work is either skipped or ends up done inline by the tier above, and nothing announces that a whole layer of the design never existed.

Why the others are wrong — Nesting is permitted rather than forbidden, so the first tier of delegation behaves exactly as intended and the design appears to work at first glance. Nor is the depth unbounded, which is what turns a three-tier orchestrator into a design fault rather than merely a slow build. And there is no re-parenting: blocked launches simply do not happen, so the missing results are absent rather than delivered to the wrong agent. Two further conditions can block a nested launch even within the limit — the current mode has to grant Task tool access, and hooks or tool policies can forbid spawning.

Remember this — Delegate two levels deep at most, and design the third tier out.

Sources — Cursor's Subagents documentation.