RL concepts, not agents
The primary thing that sets coax apart from other packages is that is designed to align with the core RL concepts, not with the high-level concept of an agent. This makes coax more modular and user-friendly for RL researchers and practitioners.
You’re in control
Other RL frameworks often hide structure that you (the RL practitioner) are interested in. Most notably, the neural network architecture of the function approximators is often hidden from you. In coax, the network architecture takes center stage. You are in charge of defining their own forward-pass function.
Another bit of structure that other RL frameworks hide from you is the main training loop. This makes it hard to take an algorithm from paper to code. The design of coax is agnostic of the details of your training loop. You are in charge of how and when you update your function approximators.