Interoperability between TF-Agents and TF-GNN

[patrik-zori]

Hello.
I wanted to ask if there is any interoperability between TF-Agents and TF-GNN (the TensorFlow library for Graph Neural Networks), both of which are official sub-projects of TF.
I am interested in training a DQN agent in TF-Agents, where the QNetwork used by the agent is a GNN built in TF-GNN. That also means that likewise, the environment/observation space will be wrapped into TF-GNN's GraphTensor class, as the environment is a graph.

If so, are there any examples? If not, how difficult would you guess hacking something functional together would take? I know I always have the option of "flattening" my graph environment into a standard Gym env and then using a standard NN for the agent, but based on the research I've read so far on the topic this would make the solution less generalizable to larger problem instances.