Your post suggests authorization as a feature:

> For each replay that we run, Roark checks if the agent follows key flows (e.g. verifying identity before sharing account details)

I don't know if AI will be more susceptible or less susceptible to phishing than humans, but this feels like a bad practice.