Forward Error Correction (FEC) has become an integral part of communication technology to address expected losses during transfer of data. Among various layer three block codes, Random Linear Network Coding (RLNC) has emerged as an adaptable, powerful approach. However, its most straightforward implementation, Full Vector Coding (FVC), introduces too much delay for widespread adoption. Handcrafted schemes were introduced to optimise coding delay, while keeping resilience reasonably high. These works have resulted in tailRLNC and PACE. For the first time we analyse their respective behaviours in a fair and comparable manner, as non-recovered packets were statistically ignored previously. We then introduce an environment that uses a consistent, unbiased simulator and interface it with a Deep Reinforcement Learning (DRL) agent. This is the first time RLNC is joined with DRL. Our deep Q-network (DQN) based agent effectively uses an optimisation loop and utilises a customisable, expressive and extendable parametric loss function to learn a protocol. We demonstrate our agent recovers hand-tailored schemes and achieves state of the art.