agree there. there's usually thinking traces involved with tool call results that become a subset of total token usage, but multiple sub-agents would easily become the majority