The entire Von Neumann architecture is not suitable for agents.

Putting data and instructions in the same memory was always a bad idea - LLMs just took this to the extreme by making data and instructions the same thing.