Effective energy resource management in power systems requires accurate prediction of wind power. In this paper, a probabilistic forecasting framework with a model-agnostic approach is presented, which combines model-agnostic meta-learning (MAML), proximal policy optimization (PPO), and adaptive quantile functions in order to improve adaptability and calibration under changing conditions. MAML allows quick adjustment to new wind farms and varying weather conditions, addressing data scarcity in new situations. PPO adjusts the quantile weighting of the training objective, supporting stable learning in dynamic scenarios. A convolutional neural network and long short-term memory (CNN-LSTM) network trained through MAML-PPO can adjust prediction strategies dynamically and estimate the parameters of adaptive quadratic spline quantile functions. The adaptively weighted continuous ranked probability score (CRPS) is confined to the training criterion, while all reported evaluations use the unweighted CRPS. Validation on two Croatian wind farms over 1-to-24-hour horizons shows aggregate accuracy statistically indistinguishable from the strongest quantile baseline on a one-week window (Diebold-Mariano p > 0.05 at horizons of 3 h and beyond), while adaptation to an unseen farm from 96 support samples (about four days) reduces CRPS by up to 16% at the one-hour horizon. On a year-long window under measurable covariate shift, static conformal calibration leaves every model miscalibrated, and trailing-window recalibration restores 86% to 91% coverage at the nominal 90% level. The framework is most valuable where data is limited or where conditions change faster than a local model can be retrained.