Token consumption in multi-agent systems is typically far higher than in single-conversation scenarios, because they involve multiple rounds of dialogue, role switching, system prompts, and context passing.