DDoS attacks plague the availability of online services today, yet like many cybersecurity problems are evolving and non-stationary. Normal and attack patterns shift as new protocols and applications are introduced, further compounded by burstiness and seasonal variation. Accordingly, it is difficult to apply machine learning-based techniques and defences in practice. Reinforcement learning (RL) may overcome this detection problem for DDoS attacks by managing and monitoring consequences; an agent's role is to learn to optimise performance criteria (which are always available) in an online manner. We advance the stateof-the-art in RL-based DDoS mitigation by introducing two agent classes designed to act on a per-flow basis, in a protocol-agnostic manner for any network topology. This is supported by an indepth investigation of feature suitability and empirical evaluation. Our results show the existence of flow features with high predictive power for different traffic classes, when used as a basis for feedback-loop-like control. We show that the new RL agent models can offer a significant increase in goodput of legitimate TCP traffic for many choices of host density.