Following a general trend in artificial intelligence, Evolutionary Computation has, in recent years, witnessed substantial performance gains from landscape-aware selection and parameter tuning of algorithmic modules. Such approaches, however, are critically relying on suitable benchmarks, or training sets, that provide the appropriate blend of performance and generality. With this position paper we argue that, on a landscape analysis basis, the benchmark design problem will form a substantial part of the next-generation of automated, on-demand algorithm design principles.