Price the debate pattern before using it: critique is ~1.2×, council is ~2.5×
Digital Applied·low signal
Multi-agent debate isn't a free quality boost. The critique variant (one model drafts, a second reviews) runs about 1.2× baseline cost; the council variant (two models in parallel with a judge arbitrating) runs about 2.5×. Reserve council-style debate for cases where the stakes justify the premium rather than reaching for it as a default reliability lever.