Recent years have seen an explosion in our ability to measure the movement dynamics of animals in naturalistic environments. However, such data poses tremendous challenges for quantitative analysis, as behavior is generally partially-observed, unpredictable, multiscale and far from equilibrium. We combine ideas from dynamical systems theory and statistical physics to draw simple yet predictive models of animal behavior across scales. We recognize and explore the interplay between theory and representation to reveal timescale separation and long-lived collective modes directly from partial observations. In addition, we introduce a fluctuating landscape picture of behavior and show how it can give rise to heavy-tailed first passage time distributions and nontrivial correlation functions.