AWS details a multimodal WhatsApp ordering assistant built on Bedrock AgentCore
A single WhatsApp Business number handles text, voice notes and calls through Amazon Nova 2 Lite and Nova 2 Sonic, sharing one backend and cross-channel memory.
AWS has published a walkthrough for deploying a multimodal WhatsApp ordering assistant built with Amazon Bedrock AgentCore and Amazon Nova 2. The target case is quick-service restaurants, where ordering is usually split across an app, a website, a phone line and the counter, leaving each channel as a separate system and fragmenting customer history. AWS notes that WhatsApp reaches more than two billion people.
One WhatsApp Business number hosts the assistant. A customer can send a text, record a voice note or start a voice call from the same conversation, and an agent takes the order from greeting to confirmation. All three channels share one backend and one cross-channel memory, so someone who texts today and calls tomorrow is recognised as the same person.
How the stack is put together
The customer front door is the Meta WhatsApp Business Platform, which exposes the Cloud API webhook plus the Messages, Media and Calling APIs. Meta operates that part; it is a prerequisite rather than something the solution deploys. Agents run on Amazon Bedrock AgentCore. Amazon Nova 2 Lite handles text via the Bedrock Converse API, while Amazon Nova 2 Sonic handles real-time speech on voice notes and calls. Agents reach the restaurant backend through the Model Context Protocol (MCP), and the whole system is deployed with the AWS CDK.
Inbound traffic lands on a single HTTPS webhook, is acknowledged with a 200 immediately and then processed asynchronously. The design separates three layers: the WhatsApp layer that handles the conversation, three agent runtimes, and a backend holding menu, carts, orders and locations. Because channel logic and ordering logic stay apart, the backend does not change when a channel is added or removed.
For voice notes, the OGG Opus bytes are decoded to 16 kHz PCM and fed into a bounded Nova 2 Sonic speech-to-speech session, with no transcription service in the path, and the reply returns as a WhatsApp voice message. For calls, Meta's Calling API delivers a WebRTC SDP offer; the runtime works in turnOnly mode because it has no public IP, TURN credentials come from Amazon KVS, and media flows over DTLS/SRTP through the KVS managed TURN relay.
Agent containers are built inside AWS CodeBuild on ARM64, so no local Python, Docker or audio toolchain is needed. Prerequisites cover both the AWS account and the Meta WhatsApp setup. The full deployment steps are in the AWS post.
