This paper presents a zero-shot adaptive elbow-assist exoskeleton controller based on deep reinforcement learning (DRL), developed to generalize robustly across variable users, tasks, and environmental conditions without per-user calibration. Trained via domain-randomized neuromusculoskeletal simulation and equipped with online latent factor estimation from surface electromyography (sEMG), the controller reduces elbow flexor effort by up to 75% in realistic lifting tasks. Human-subject experiments reveal a trade-off between muscular unloading and trajectory tracking accuracy, particularly at higher speeds and amplitudes. Despite this, the controller reliably delivers user-adaptive assistance with no subject-specific tuning. This makes it highly suited for time-sensitive, repetitive tasks in automation contexts such as warehouses, logistics, and manufacturing. In these environments, minimizing setup time and operator fatigue are critical for productivity.