AWS's phone-ordering host is really an MCP wiring diagram with no retry button

The seam worth studying is where the agent stops talking and starts calling code. AWS puts MCP there, with AgentCore Gateway publishing the restaurant's APIs as tools the agent can discover and call by name ^[7], registered for the agent's use through Amazon AppIntegrations ^[8], and fronted by API Gateway REST endpoints with IAM on them ^[9].

Follow one question through that. "Is the pad thai spicy" arrives as audio off the phone network ^[3], gets recognised by Agentic Voice ^[5], becomes a turn for the orchestration agent running Claude Haiku 4.5 ^[6], which calls a gateway tool ^[7], which crosses API Gateway ^[9] to the backend that actually holds the menu ^[10], and the answer returns as synthesised speech ^[5]. That is five legs each way inside one spoken pause ^[16]. In a browser you would put a spinner there. On a call the customer hears silence and starts deciding whether the line dropped.

The only tuning knob the walkthrough names for that problem is confidence-based end-of-turn detection in the ASR ^[5]. It is a real lever and a two-sided one: set the confidence bar high and the host waits politely through dead air, set it low and it cuts across people who were mid-sentence. AWS says a later section explains how Agentic Voice speeds up turn-taking ^[15]; the material in front of me stops before any figures. In a telephony design, that is the load-bearing claim arriving unquantified.

On modularity, the post keeps agent logic and backend as separate modules so the ordering logic does not depend on the channel calling it ^[4], and argues that because MCP is an open standard the backend can change without touching the agent ^[11]. True at the wire, and incomplete. Voice writes requirements back into the backend regardless: every field a tool returns is going to be read out loud, so a chatty response payload costs both time and comprehension, and callers repeat themselves, so a tool that adds an item to a cart has to survive being called twice.

The build split is stated plainly, and it is the most useful paragraph in the post for anyone doing a budget. Connect provisions the contact flow, the speech layer and the AI agent in a single deployment ^[12], which leaves seven of the eight named components managed and the gateway plus the backend behind it as yours ^[18]. Those are also the pieces holding the menu, the carts, the orders and the locations ^[10]. The guardrails on offer are content filters, denied topics and profanity filtering ^[13], which police what the conversation is about, not whether the order read back to the caller matches what is in the cart. Nothing else in the component list does that either, so the backend does it or nobody does.

CDK gets the whole thing stood up ^[14], and what stands up is a phone number that answers and takes orders through to confirmation ^[1]. The demo is cheap. The failure mode is a stranger on hold.