Hey how about we go back to LLM/AI being a simple knowledge request service? You have a model trained on data, you request a text or media response from that knowledge model.

No agents. And if someone loops the response into an agent, well, then they are legally liable for personally doing that.