Multi-agent debate — full mesh vs sparse star
full mesh everyone reads everyone
alpha
beta
gamma
N=3, R=2 12 critique ops
sparse star hub-and-spoke
hub
spoke1
spoke2
spoke3
spoke4
N=5, R=2 16 critique ops (vs 40 full mesh)
when debate helps vs when it hurts
HELPS: factuality (cross-check), rule-following (chess), open-ended reasoning
HURTS: latency-sensitive UX, cost-bound flows, simple factual lookup
Society of Minds: cross-model (ChatGPT + Bard) beats either alone
sparse topology: matches accuracy at fraction of token cost
reach for debate when single-shot is wrong > 5% of the time; otherwise cheaper patterns win.