My concern would be the potential for an alignment issue. The agent directs it's eyes to look at you - because it's communicating something accurate about the plan it's formed about how to cross the intersection? Or because it's been trained to look at people when it crosses intersections, but communicating isn't it's true objective?
(I'm not trying to imply the agent may fool you into crossing so it can mow you down, the case of "AI incompetence" so to speak is more concerning to me then "AI malice".)
(I'm not trying to imply the agent may fool you into crossing so it can mow you down, the case of "AI incompetence" so to speak is more concerning to me then "AI malice".)