A builder-focused guide to preventing prompt injection in AI agents. Learn why prompt engineering alone fails and how five architectural defense layers — input boundaries, least-privilege tooling, boundary authorization, human-in-the-loop gates, and tamper-evident logging — contain real attacks.