Trust boundary is the line in a system’s architecture where data or instructions pass between components with different levels of trust, so that anything crossing it must be validated, tagged, or treated as untrusted. In LLM agent design it matters because models cannot reliably tell instructions from data, so systems mark external content (web pages, tool outputs, user files) as untrusted and limit what it can make the agent do. The boundary is easy to breach by accident: when a cheap sub-agent reads untrusted material and passes a summary to the main agent, that report arrives in the system’s own voice and skips untrusted-input tagging, so injected instructions can be laundered into trusted context unless the sub-agent hand-off is itself treated as a boundary.