<p>Biomolecular condensates formed by liquid–liquid phase separation (LLPS) have emerged as fundamental components of intracellular organization. While their biophysical mechanisms are increasingly well understood, much less is known about how the ability to phase-separate is subject to natural selection. In this article, we argue that phase separation should be treated as an evolvable molecular phenotype. We review how classical sequence-based molecular evolution, motif-level modeling, and comparative trait-evolution frameworks collectively offer powerful tools for dissecting how natural selection shapes the condensate-forming capacity of proteins, and we outline promising directions for future investigation. Using examples from well-aligned proteins such as members of the FET family and from motif-rich intrinsically disordered regions (IDRs), we illustrate how selection acts on residues, motif number and spacing, domain architecture, and emergent biophysical properties. Recent advances in artificial intelligence make it possible to predict latent molecular traits—including disorder propensity, interaction valency, and phase-separation potential—directly from sequence, enabling these properties to be embedded within evolutionary models. Together, these approaches lay the foundation for the emerging field of <i>evolutionary condensate biology</i>, spanning residue-level constraints, motif organization, and biophysical trait dynamics. This synthesis opens new opportunities for comparative and predictive studies of LLPS across the tree of life and for the rational design of proteins capable of controlled phase separation.</p>