How many agents does a typical production multi-agent system contain?	3–7 agents.
What is the token-cost multiplier of a three-agent debate versus the best single agent on the same task?	~5× more tokens.
What accuracy lift does debate typically produce on hard tasks?	~10–25% over the best single agent.
Name the three walls that single-agent systems hit, as listed in the chapter.	Context length, specialization, and error correction.
What is the 'collusion' failure mode of the debate pattern?	Two instances of the same model will often agree on the same wrong answer, especially when the model is overconfident — they fail to genuinely challenge each other.
How many pairwise interfaces does a five-agent system have?	Ten — C(5,2) = 10.
What does the chapter say is the most under-discussed cost of multi-agent systems?	Debugging: tracing a failure across ten pairwise interfaces is harder by an order of magnitude than debugging a single agent.
What communication pattern does the chapter say rarely survives in production?	Plain text turns in a shared transcript — a research convenience that becomes wasteful and impractical past three or four agents.
What three things does the chapter say multi-agent systems actually buy you?	Specialization, error correction, and context isolation.
In what domains does the chapter say the debate pattern's accuracy lift is typically worth the 5× cost?	High-stakes settings: legal review, medical synthesis, financial decision support.