<p>High-entropy alloys exhibit exceptional performance under extreme environments; however, conventional equiatomic or near-equiatomic design paradigms restrict exploration of single-phase solid-solution spaces, leaving many non-equiatomic compositions unexplored. Here, we propose a knowledge-enhanced AI framework integrating physics-guided machine learning with large language model (LLM) agents to explore compositions beyond equiatomic constraints. Machine learning models trained on physics-based descriptors achieve F1-scores of 0.96 and 0.94 for face-centered cubic (FCC) and body-centered cubic (BCC) phase prediction, respectively. Two substantially non-equiatomic alloys—Al₂₈Cr₂₂Fe₂₆Ni₁₅Mo₉ (BCC) and Co₂₃Cr₁₅Fe₂₅Ni₃₁Mo₆ (FCC)—experimentally confirm reliable phase prediction within the investigated compositional domain. LLM-guided reasoning on oxidation mechanisms further enables the design of Co₂₂Cr₂₄Fe₂₀Ni₂₅Mo₅Mn₄, which exhibits a steady-state oxidation rate constant of 1.41 × 10⁻⁸ mg^<i>n</i> cm^−2<i>n</i> h⁻¹ (50–100 h, <i>n</i> = 45.35) at 900 °C, indicative of self-limiting oxidation kinetics. This superior performance is attributed to the formation of a stable Cr₂O₃/(Mn,Fe)Cr₂O₄ bilayer oxide scale operating via a synergistic barrier–buffer protection mechanism. This study presents a data-efficient, AI-assisted methodology for intelligent HEA design in high-dimensional compositional spaces.</p>