🔗 The agent needs identity. The model does not.

A support agent may need a customer's email address to perform a lookup, but importantly, the model does not require that information. This article delves into the challenges of maintaining this distinction throughout an agent's operations, especially as it starts calling tools. With a practical demonstration using a support agent built with Phoenix LiveView, OpenAI, and other tools, it explores how personal information can inadvertently escape through various application states and tool arguments. The author outlines strategies to protect identity relationships while allowing necessary data access for operations. By designing these safeguards, agents can operate effectively without exposing sensitive customer information to models. This ensures a balance between functionality and privacy, showcasing a modern architectural approach to identity management in applications.