Neither, probably. The more likely outcome is that the definition of "coding" shifts. The people who understand what the code should do — the system design, the edge cases, the user problem — will remain essential. What gets automated is the mechanical translation of that understanding into syntax. That's actually a good thing for software quality if the people directing the AI know what they're doing.