Tensorflow how to train a model step by step?

Question

Neural networks

Many demo videos on neural networks show the sequential process of training the network, they clearly show how, with each new iteration, the model gradually begins to make more or less expected decisions.
I'm studying the topic, I'm interested, I've already written several models that work, do what is required of them: one RNN model with the task of text classification, the second is based on Keras DQNAgent - it plays a simple toy.
Each of these models was trained in its own way: 3 million texts were prepared for the first, the second just played its own game for 50,000 iterations.

Now the question is: how to train the network exactly in steps? Not doing model.fit on a huge dataset, or leaving it for 50k "self-education" cycles and doing each step directly with a human?
Let for example it will be DQNAgent. A new pong game cycle is launched and the opponent of this network is a human, and in the process of playing with a human, the network must learn.