Today's networks are struggling to scale and satisfy the high number and high variety of co-existing network requirements. While existing congestion control (CC) protocols are designed to handle strict classification of network flows into one or few priorities, a more granular and dynamic congestion control is needed.In this paper we present Hercules, a novel CC protocol based on an online learning approach, which supports unbounded and continues requirements space. We implemented Hercules as a QUIC module and we show, through analytical analysis and realworld experiments, that it provides between 50% − 250% higher QoS for co-existing diverse network flows and outperforms stateof-the-art CC protocols, even under high network congestion.