It's not really accurate to say it's "only considering one token at a time". Foresight and (implicit) planning are taking place. You can see this clearly during programming tasks, where imports come hundreds of tokens before they are eventually used.