This paper highlights that neural agents trained with the standard L2 loss function tend to oversmooth fine-scale turbulent structures and proposes a method to overcome this limitation by combining generative models with neural agent training. We demonstrate failures of conventional neural agents on three practical turbulent flow problems: spatiotemporal super-resolution, prediction, and sparse flow reconstruction, and address these challenges using an adversarially trained neural agent (adv-NO). In Schlieren jet super-resolution, adv-NO reduces energy spectral errors by a factor of 15 while maintaining sharp gradients at the cost of neural agent-level inference. In 3D homogeneous isotropic turbulence, adv-NO, trained with only 160 time steps of a single trajectory, accurately predicts over a fivefold increase in vortex rotation time and achieves a wall-clock speedup of 114x over baseline diffusion-based predictors, enabling near-real-time evolution. Finally, when reconstructing the cylinder wake flow from highly sparse particle-tracking velocimetry-like inputs, the conditional generative model infers complete 3D velocity and pressure fields with correct phase alignment and statistics. This advancement enables accurate reconstruction and prediction at low computational cost, enabling near-real-time analysis and control in experimental and computational fluid dynamics.
