A novel adaptive compensation bipartite containment control strategy based on improved dynamic surface control (DSC) method and radial basis function neural networks (RBFNNs) is designed for nonlinear stochastic multi-agent systems (SMASs). A definition for implementing containment control in probability is proposed. A high-gain observer is employed to reconstruct the unmeasurable system state, and the non-strict feedback term in the unknown function is reduced by a new processing method. A second-order command filter (CF) is constructed to eliminate the influence of filtering error on tracking performance and simplify the complexity of the controller.