Multi‑Agent Debate (MAD) improves the reasoning accuracy of large language models (LLMs) through iterative peer interaction, and the communication topology is central to this process. Consequently, researchers have proposed increasingly sophisticated mechanisms—learning, adapting, or dynamically reconfiguring agent interactions—to boost accuracy or reliability.\ \ Prior studies, however, indicate that very simple sparse communication can already achieve competitive performance at a much lower cost.\ \ In this work we focus on sparse MAD and introduce an ultra‑simple random‑without‑replacement routing policy: at each round every agent randomly selects two distinct peers that it has not yet debated with and engages in a debate. The method requires no learning or extra scheduling, relying solely on random sampling to generate a diverse interaction graph.\ \ Empirical results show that this Distinct‑Peer random routing consistently improves the accuracy‑cost trade‑off, establishing a surprisingly strong baseline for sparse MAD.\ \ Building on this observation, we study lightweight deliberation stopping. Simple stopping criteria can substantially cut inference cost while preserving competitive accuracy.\ \ Overall, unless a more complex learned or adaptive topology can clearly outperform these simple baselines, its added complexity is not justified.\ \ Review