I agree with your point.

But I also find interesting that the agent wasn't instructed to write the hit piece. That was on its own initiative.

I read through the SOUL.md and it didn't have anything nefarious in there. Sure it could have been more carefully worded, but it didn't instruct the agent to attack people.

To me this exemplifies how delicate it will be to keep agents on the straight and narrow and how easily they can go of the rails if you have someone who isn't necessarily a "bad actor" but who just doesn't care enough to ensure they act in a socially acceptable way.

Ultimately I think there will be requirements for agents to identify their user when acting on their behalf.