Week 0: Your Agent Passes in Text and Fails on a Real Call
Prerequisite: none. This is a short framing lesson before Week 1, not a technical lesson. After this lesson, you can: name why a voice agent that reads perfectly in a chat transcript can still fail on a real call, and know whether that gap is currently invisible to your own testing.
If you've tested your agent by reading its transcripts, you've tested the wrong thing. A transcript shows you what got said. It doesn't show you who was allowed to interrupt whom, whether a pause meant "finished" or "still thinking," or what happened when three concurrent systems disagreed about what the caller had just said. Those are voice problems specifically, and text testing is structurally blind to every one of them.
The mistake this course corrects
If your instinct is that voice-agent quality is now a component-quality problem (better transcription, more natural speech, a faster model), that instinct made sense a few years ago and doesn't anymore. Lesson 02 lays out exactly where the industry stands in 2026 and names its sources; the short version is that ASR, TTS, and raw component latency have mostly stopped being the bottleneck. What's unsolved is the seams between them under real conversational pressure: turn-taking, cancellation, state, and the compliance gates a live call has to pass that a text thread never does. This course names six specific ways that shows up, starting in the next lesson with two numbers from published turn-taking research, cited by name there.
What you'll be able to do that you can't do now
You can build a voice pipeline that sounds right in a demo call and can't explain why it feels wrong on a real one.
You can name the specific orchestration gap behind a failure (turn-taking, backchannel misclassification, split state, premature goodbye, spoken hallucination, or handoff amnesia) and apply one of five reusable patterns that fixes it, without waiting on a smarter model.
Concretely, you'll build a two-tier classifier that correctly reads the same word as an interruption or a backchannel depending on context, design cooperative cancellation so a caller's "stop" actually stops every in-flight operation, and leave knowing the two regulatory regimes and the voice-specific metrics that predict abandonment before a customer ever complains.
Why this is architecture, not model quality
Every one of the six failures above has the same shape: a component did its individual job correctly, and the system still broke because of how the pieces were wired together under real timing pressure. That's not a coincidence, and it's not a case where waiting for next year's model helps. A faster, smarter LLM doesn't decide who's allowed to interrupt whom, or what "stop" should cancel. Only the orchestration layer does, and that's the layer this course actually teaches.
Who this is for
Comfort with the idea of an LLM calling tools is assumed; this course doesn't re-teach tool-calling mechanics from scratch. No prior voice-agent or telephony experience is required: ASR, TTS, and SIP are each introduced from the role they play in the failure they cause, not from first principles.
Two measured numbers from turn-taking research, the contradiction they create, and why a faster model can't close the gap on its own.
Reply here and it goes straight to Rod. Same as replying to one of his emails.
