We develop a general theory of agnostic online learning from continuous-time data streams under limited queries, providing tight regret bounds for both oblivious and adaptive settings.