The agent prompt is stored as two parts:
- Prompt — the stable base written by engineering or product. SManager-only to protect core agent behavior.
- Extra Instructions — brand-specific or campaign-specific additions that operators can adjust without touching the base prompt. Any user with the Edit role can update this field.
When the agent is invoked, both sections are joined into one combined prompt with a --- separator between them. When you load an existing agent, the editor splits the stored prompt back into these two fields automatically.
This separation means you can safely give content or customer-success teams the ability to tune agent behavior (Extra Instructions) without risking accidental changes to the core prompt logic.