Simulation-based optimization using agent-based models is typically carried\nout under the assumption that the gradient describing the sensitivity of the\nsimulation output to the input cannot be evaluated directly. To still apply\ngradient-based optimization methods, which efficiently steer the optimization\ntowards a local optimum, gradient estimation methods can be employed. However,\nmany simulation runs are needed to obtain accurate estimates if the input\ndimension is large. Automatic differentiation (AD) is a family of techniques to\ncompute gradients of general programs directly. Here, we explore the use of AD\nin the context of time-driven agent-based simulations. By substituting common\ndiscrete model elements such as conditional branching with smooth\napproximations, we obtain gradient information across discontinuities in the\nmodel logic. On the example of microscopic traffic models and an epidemics\nmodel, we study the fidelity and overhead of the differentiable models, as well\nas the convergence speed and solution quality achieved by gradient-based\noptimization compared to gradient-free methods. In traffic signal timing\noptimization problems with high input dimension, the gradient-based methods\nexhibit substantially superior performance. Finally, we demonstrate that the\napproach enables gradient-based training of neural network-controlled\nsimulation entities embedded in the model logic.\n