It’s still going to chew up context quickly. Surely, some of the added tokens are helping the model, but does it require as many as it generates? What happens on long, multi step tasks as it pushes old tokens out of context? I’m not sure we know the answers to those.