This paper presents an original multi-level optimisation method for the design of composite structures integrating a global–local approach based on higher-order theories to assess the responses of the structure at each scale. The method offers a good balance between accuracy and computational costs. Unlike multi-level strategies available in the literature, in the proposed approach there is a strong interaction between the steps of the optimisation process. The proposed method is articulated in two nested optimisation loops (outer and inner). The outer loop focuses on the macroscopic scale where the polar formalism is used to describe the laminate behaviour. The resolution of the outer loop is performed through a special metaheuristic algorithm. However, since requirements on local structural responses are evaluated on the most critical region of the structure (modelled through a higher-order theory) at the ply-level, for each solution of the outer loop, a nested genetic optimisation (inner loop) is performed to find the stack matching the values of the geometric variables and of the polar parameters corresponding to the current solution of the outer loop. During the inner loop, the optimised stacking sequences are searched in the domain of general quasi-trivial solutions, without introducing simplifying hypotheses. The new methodology is applied to the least-weight design of a simplified wing-box structure by considering requirements of both mechanical nature (first buckling load, first-ply failure, and delamination) and technological nature (blending between adjacent laminates).