The AI voice agent space is drowning in noise. 10,000+ LinkedIn posts a month, almost none showing architecture. This talk draws a hard line between demo-grade and carrier-grade.I'll walk through SignalWire's three-layer AI Agent control plane: typed functions, state machines, and prompts, and why collapsing them into a single system prompt is the root cause of every production failure nobody posts about. I'll introduce System Directed AI, the architectural principle that separates code-driven business decisions from AI-driven conversational decisions, and show a concrete System Directed AI technique: zero-argument tool calls that read from validated stored state instead of letting the LLM re-supply data on every call. Fewer arguments, fewer failures, guaranteed consistency.Everything shown is open source at https://github.com/signalwire-demos. No demos, no happy-path videos. Just architecture, code, and the failure modes the industry skips.