Planning requires an agent to anticipate the consequences of its actions and reason about how the state of the world may change. Classical approaches to planning rely on explicit states, actions, and transition dynamics. They're often expressed symbolically. Recent advances in representation learning offer the possibility of learning world models directly from high-dimensional observations, without requiring an explicit symbolic description of the state of the environment/world. This forces us to confront a foundational question about the role of explicit symbolic states for planning vs sufficiency of learned representations. This work investigates the use of representation-learning-based world models for planning. In particular, I study encoder-based architectures such as Joint-Embedding Predictive Architectures (JEPAs). JEPA based world models learn by predicting the latent representations of future states rather than reconstructing observations at the pixel level. Such architectures provide a attractive foundation for planning because they aim to capture predictable structure while abstracting away irrelevant details of the observation space. I train and evaluate these models on a range of planning tasks. Rather than evaluating world models through representation quality or predictive accuracy, the central criterion is their utility for planning. I investigate whether the representations and latent dynamics learned by these models capture the underlying dynamics sufficiently well. This includes studying how errors in learned latent dynamics accumulate and identifying the properties of the environment important for planning.