You need to be able to decompose tasks into smaller pieces, such that it can be solved by a set of agents none of whom see the whole task (and each of those pieces needs to be decomposable…). We use the decomposition to produce training data, but in the end we train an agent who solves the whole task.