Guarding Graph Boundaries

The boundary is where the user meets the AI. This is the danger zone.

The Sandwich Pattern

Wrap your core logic in guards.

graph TD
    Input --> InputGuard{Safe?}
    InputGuard -- No --> Block[Block Request]
    InputGuard -- Yes --> CoreAgent[Core Reasoning]
    CoreAgent --> OutputGuard{Safe Output?}
    OutputGuard -- Yes --> User
    OutputGuard -- No --> Rewrite[Rewrite Logic]
    Rewrite --> OutputGuard

Input Guards (Pre-Processing)

Before the LLM even sees the message, check for PII or Prompt Injection using a cheap model or regex.

Output Guards (Post-Processing)

Before showing the answer, ensure the agent didn't "leak" keys or generate harmful content.