This sounds like they have an LLM running with a context window that just gets longer and longer and contains all the past interactions of the store.
The normal way you'd build something like this is to have a way to store the state and have an LLM in the loop that makes a decision on what to do next based on the state. (With a fresh call to an LLM each time and no accumulating context)
If I understand correctly this is an experiment to see what happens in the long context approach, which is interesting but not super practical as it's knows that LLMs will have a harder time at this. Point being, I wouldn't extrapolate this to how a commercial system built properly to do something similar would perform.
replies(2):