Aviation faces increasing pressure to reduce emissions and environmental impact, requiring more efficient lightweight aircraft. Improving aerodynamic efficiency generally tends to generate elongated wings characterised by heavy airframes. Wing mass itself must, therefore, be reduced through dedicated structural optimisation to maintain overall performance. Achieving this balance demands high-fidelity structural models remaining affordable within gradient-based optimisation loops. This paper presents a computationally efficient gradient-based structural optimisation framework for composite wing structures combining high-order one-dimensional beam models with adjoint sensitivities and automatic differentiation. Structural models rely on a 1D finite element framework based on the Carrera Unified Formulation, where cross-sectional kinematics are represented through hierarchical expansion functions. In particular, Taylor expansions are used to generate higher-order equivalent beam models, while Lagrange expansions provide a layer-wise description of the cross-section. The framework is used to model a composite wing box with panels stiffened by Z-stringers, representative of a modern aircraft wing bay. The wing-box mass minimisation is formulated by treating the thicknesses of skins and stiffeners as design variables, while allowable strain constraints are aggregated through the Kreisselmeier–Steinhauser function. Exact gradients are obtained from a discrete adjoint method, implemented using reverse-mode automatic differentiation and supplied to a Sequential Quadratic Programming solver. A third-order Taylor expansion is selected for the optimisation, based on prior assessments proving that this order yields strains and stresses comparable to commercial 2D and 3D finite element models while drastically reducing computational cost. Results confirm that adjoint-based gradients combined with constraint aggregation enable efficient optimisation with several design variables, achieving mass reductions comparable to industrial shell-based workflows at markedly lower computational expense.