80% of multi-agent performance variance is just token allocation, not the model. That's from a UC Berkeley paper that decomposes any AI agent into 6 components: model, memory, context, skills, orchestration, verification. The model is one. The other 5 are where the gains