Deep Learning (DL) has achieved automatic and objective assessment of surgical skills. However, the applicability of DL models is often hampered by their substantial data requirements and confinement to specific training domains. This prevents them from transitioning to new tasks with scarce data. Therefore, domain adaptation emerges as a critical element for the practical implementation of DL in real-world scenarios. Herein, we introduce A-VBANet, a novel meta-learning model capable of delivering domain-agnostic surgical skill classification via one-shot learning. A-VBANet has been rigorously developed and tested on five diverse laparoscopic and robotic surgical simulators. Furthermore, we extend its validation to operating room (OR) videos of laparoscopic cholecystectomy. Our model successfully adapts with accuracies up to 99.5% in one-shot and 99.9% in few-shot settings for simulated tasks and 89.7% for laparoscopic cholecystectomy. This research marks the first instance of a domain-agnostic methodology for surgical skill assessment, paving the way for more precise and accessible training evaluation across diverse high-stakes environments such as real-life surgery where data is scarce.