Put another way, ultimately what's the distinction between a prompt and an index to stored data?
If you have to actually account for all the working details of Minecraft in the prompt to get the agent to express your intent, isn't the code for Minecraft the most certain way to approach this?
A program is a precise behavioral specification. In what ways can this be simplified by "prompting" except to ask these agent to "emit the Minecraft program", but if the agent had never seen the Minecraft code, then how would it determine what to emit?
Was it "shown" a working form of Minecraft? Who made sure the demo was sufficient enough to account for completeness of Minecraft? How was this completeness expressed? If the agent can't actually produce Minecraft and emits an app that's Minecraft-like, then isn't the titular claim false?
If the model has been exposed to the code for literal Minecraft, then how is provenance of the original code accounted for in the agent's output, which amounts to a clever kind of retrieval of previously input data?
It seems the least interesting aspects of this article's claims have anything to do with Minecraft. A most interesting question is why and how is it easier to create a prompt to produce an app than to program the app yourself?