Multiagent systems have received considerable attention from many disciplines, including biology, physics, engineering, and economics, and novel applications of these systems lead to interesting challenges in their modelling, analysis, and control. In particular, large-scale multiagent systems require the use of techniques from kinetic theory, with the mean field limit as a prime example. Furthermore, considering complex multiagent systems as arising in biology and sociology, for instance, makes the use of first-principles modelling challenging, which motivates the use of data-driven approaches and machine learning instead. In this context, we establish new connections between multiagent systems, kinetic theory, and kernel methods in machine learning. In order to perform machine learning on large-scale multiagent systems, we consider learning with kernels in the mean field limit. This requires a suitable hypothesis class, which we provide by kernels in the mean field limit, and we show that the resulting reproducing kernel Hilbert spaces (RKHSs) are compatible with the mean field limit. Kernel methods built upon this are then suitable for statistical learning problems in the mean field limit. In particular, this allows to move between the microscopic and mesoscopic level, which opens up the possibility of learning on very large-scale multiagent systems. Our work leads not only to new connections between kernel methods and kinetic theory, but also provides solid, rigorous foundation for novel learning methods in the context of complex, large-scale multiagent systems.