AWS details multimodal WhatsApp ordering assistant built with Bedrock AgentCore

AWS has published a technical guide for deploying a multimodal restaurant-ordering assistant that operates through a single WhatsApp Business number. Built with Amazon Bedrock AgentCore and Amazon Nova 2, the assistant can receive text messages, voice notes and voice calls.

AWS says an AI agent manages the ordering conversation from the initial greeting through confirmation. The three channels use the same backend and cross-channel memory, allowing the system to recognize a customer who uses text during one interaction and calls during another.

Models and agent runtimes

The design uses the Meta WhatsApp Business Platform as its customer-facing layer. Amazon Bedrock AgentCore hosts three agents, with separate runtimes for the supported channels. According to AWS, each conversation runs inside its own microVM and sessions remain isolated.

Amazon Nova 2 Lite processes text through the Amazon Bedrock Converse API. Amazon Nova 2 Sonic handles real-time speech for voice notes and calls. The agents communicate with the restaurant backend through the Model Context Protocol, or MCP.

AWS separates the system into three parts: the WhatsApp conversation layer, the agent runtimes and the restaurant backend. The backend maintains the menu, carts, orders and locations. AWS says separating channel handling from ordering logic allows channels to be added or removed without changing the backend.

Webhook and backend architecture

Incoming WhatsApp traffic reaches a single regional HTTPS webhook. The webhook immediately returns an HTTP 200 response, after which the request is processed asynchronously. The architecture uses Amazon Simple Queue Service to separate the initial acknowledgment from subsequent processing. The queue includes a dead-letter queue.

Amazon API Gateway supplies two REST APIs. One is the regional HTTPS webhook, which uses an AWS-managed certificate and serves as the system’s only public endpoint. The other is an IAM-authorized API placed in front of the ordering backend.

AWS Lambda runs the webhook ingestion function, webhook worker, message sender and ordering business logic. The Meta platform provides the Cloud API webhook, Messages API, Media API and Calling API used by the customer-facing portion of the system.

Deployment responsibilities

The guide uses AWS Cloud Development Kit to provision the AWS components as multiple stacks deployed in dependency order. The resulting deployment provides a webhook URL that must be registered with Meta.

The AWS CDK deployment does not create or manage the Meta WhatsApp Business Platform. Users must set up that service and connect it to the AWS deployment as a prerequisite. AWS describes the Meta platform as the assistant’s front door and the AWS environment as the host for its agents, processing components and ordering backend.

Source: Amazon Web Services’ “Deploy a multimodal WhatsApp ordering assistant with Amazon Bedrock AgentCore,” published September 4, 2026.