Immunology has made great headway in mapping the complexity of immune responses, yet it lacks a framework to derive theoretical understanding from high-dimensional datasets. We combined a robotics platform with machine learning to build a simple model of cytokine kinetics produced by mouse CD8+ T cells in response to antigens of diverse potencies (or qualities). We found that the high-dimensional cytokine dynamics can be projected onto a low-dimensional latent space in an antigen-specific manner. We defined this projection as “antigen encoding” and derived dynamical equations describing trajectories in this latent space. Our latent space model was critical in uncovering new immune patterns following chimeric antigen receptor (CAR)-T cell activation. Such modelling of T cell activation can inspire similar latent space approaches to derive simple dynamical models from high-dimensional biological datasets.