"Who bears legal responsibility when an agent blows past its guardrails and causes real-world harm?"
That is very easy to answer - the person being responsible for what is running on that hardware when it's personal hardware, and in corporate environments the manager who is responsible for deploying the agent and the engineers who didn't ensure this doesn't happen if they failed in their responsibilities by not monitoring the system (dependent on if the manager even allowed suffcient ressources for monitoring, if not, they should also not share the blame)