The distinction I’d add is that AI doesn’t necessarily make engineers more replaceable undifferentiated output does.

If an agent can generate the implementation, the durable value shifts toward knowing what should be built, spotting when the generated solution is wrong, debugging across system boundaries, and understanding the trade-offs behind the architecture.

The “AI dependency score” is a useful idea, but I’d measure something slightly different: how much of the system can you still explain, debug, and change without the agent. If that number keeps shrinking while your output rises, the productivity gain is probably borrowing from your future skill set.