Deploy a Multimodal WhatsApp Ordering Assistant with Amazon Bedrock AgentCore

This article details the deployment of a multimodal WhatsApp ordering assistant leveraging Amazon Bedrock AgentCore and Amazon Nova 2. The system enables customers to place orders via text, voice notes, or voice calls through a single WhatsApp Business number, providing a seamless, cross-channel experience. The assistant recognizes returning customers across all channels using shared memory, ensuring continuity and personalized interactions.

The architecture separates the conversation layer (handled by Meta WhatsApp Business Platform) from the processing layer, which includes three agent runtimes hosted on Amazon Bedrock AgentCore. Amazon Nova 2 Lite manages text interactions, while Amazon Nova 2 Sonic handles real-time speech for voice notes and calls. These runtimes communicate with the backend via the Model Context Protocol (MCP), which exposes REST APIs for menu retrieval, order management, and customer data.

Deployment is managed through the AWS Cloud Development Kit (AWS CDK), which provisions resources such as API Gateway, Lambda functions, Amazon SQS queues, DynamoDB, and networking components. The system ensures fast acknowledgment of incoming messages by immediately responding to Meta webhooks, while processing the rest asynchronously through a queue and worker pattern. This design maintains high responsiveness and reliability.

The shared memory, implemented via Amazon DynamoDB, maintains customer profiles, order history, and preferences, enabling the assistant to recognize customers across channels without storing personally identifiable information (PII) directly. The system supports order placement, menu browsing, and location-based services, with potential for expansion to handle images and documents.

Prerequisites include setting up the Meta WhatsApp Business Platform, obtaining necessary tokens and IDs, and configuring AWS resources with appropriate permissions. Deployment involves cloning the sample repository, running setup scripts, and completing webhook registration and verification steps. The entire process is designed to be repeatable and adaptable for different business needs.

This solution is suitable for retail, healthcare, field service, and other domains requiring multimodal, cross-channel customer interactions. It emphasizes security, scalability, and ease of integration, with options to enhance content filtering and policy compliance using Amazon Bedrock Guardrails. Proper cleanup procedures are provided to decommission resources when no longer needed.