Abstract, or disentangled, representations are a promising mathematical framework for efficient and effective out-of-distribution (OOD) generalization. Reflecting the topology of the real world in their representational geometry, they offer a compelling reason for why both biological and artificial systems might converge to such interpretable representations. We here highlight recent results demonstrating that disentanglement comes about naturally when (biological or artificial) agents solve canonical decision-making tasks that require evidence aggregation, in parallel. The tasks tie closely to Bayesian filtering theory, and should be solved by any agent that deals with a noisy world.Intriguingly, theory and experiments together suggest that solving such day-to-day decisions involving latent variables present in the real world directly leads to representations that (1) preserve the topology of the world, (2) isolate such factors of variation, and (3) are consistent across individuals. Furthermore, the highlighted work provides exact mathematical conditions for the emergence of such representations, and demonstrate the importance of noise in facilitating disentangling. The findings are consistent across tasks types and architectures, and we find that transformers are particularly suited for disentangling representations, which might explain their unique world understanding abilities. The universality of the tasks makes us believe that they present a prime candidate for OOD generalization in the brain. Hence, it should be no surprise that such disentangled, topology-preserving representations are widely found in the brain, in examples as disparate as navigation, decision-making and memory. We here expand upon and discuss in detail about potential implications of these findings, for machine learning and neuroscience alike.