As a healthcare company, we are currently going through the process of how to allow on-device AI activity while reducing the associated risk. I really like the idea of giving the agent a specific identity instead of globbing onto the credentials of the person operating it. I think I would like to go further and splitting the agent out onto its own hardware complete separate from the developer. Sounds extreme but MacMini devices are pretty cheap and fairly capable.