also supports subagents through parallel mid turn task delegation on multiple gpu slots, fanning out on available and queuing tasks while busy