Debate helps most when the agents genuinely disagree for a reason, not when they share the same model, prompt and blind spots. If three copies of the same model argue, they mostly converge on the same error, politely. The gain appears when the agents differ in a meaningful way: different tools, different retrieved evidence, different roles, or different models. Then a disagreement is a signal that something is underspecified or uncertain, and the debate can surface it.
The practical design is to give debate a narrow job. Use it for verification, not generation: let one agent produce an answer with reasoning and evidence, and let the others attack specific claims, check calculations, or look for missing cases. A judge that only sees the final arguments, rather than the full transcript, also helps, because long debates tend to reward confident phrasing over correct reasoning.
On cost, the rule I use is simple: debate only when the task is hard enough that a single pass is unreliable, and cheap enough that a second opinion is worth more than a retry. For routine questions, self-consistency — sampling several independent answers and taking the majority — usually gives most of the benefit at lower cost. Reserve full debate for high-stakes decisions, ambiguous requirements, or cases where you need an audit trail for why an answer was chosen.