Dr. Sean Kelly of Imprivata recommended that AI agents have clearly defined identities, permissions, and boundaries, with human review for higher-risk activities and audit trails.