AWS Publishes Reference Architecture for Multimodal WhatsApp Ordering Agents Using Bedrock AgentCore and Nova 2

What AWS Announced

AWS has published a detailed reference architecture for deploying a multimodal ordering assistant on WhatsApp, built with Amazon Bedrock AgentCore and Amazon Nova 2 models. The post, aimed at quick-service restaurant operators, shows how to run text, voice note, and voice call ordering through a single WhatsApp Business number backed by one shared customer memory.

This is a deployment guide and architecture pattern, not a new model release. The underlying AI components — Amazon Nova 2 Lite and Amazon Nova 2 Sonic — are invoked through existing Amazon Bedrock APIs.

The Technical Setup

The architecture separates three layers: the WhatsApp conversation layer, three per-channel agent runtimes, and a backend holding menu, cart, and order data. Each conversation runs in an isolated microVM on AgentCore runtime.

Key components:

• Amazon Nova 2 Lite handles text conversations through the Bedrock Converse API
• Amazon Nova 2 Sonic handles real-time speech-to-speech for voice notes and voice calls
• AgentCore Gateway exposes the restaurant backend REST API as MCP (Model Context Protocol) tools — including GetMenu, AddToCart, and PlaceOrder — that agents call by name
• AgentCore memory maintains one shared, cross-channel record keyed by a hashed customer ID, so a customer who texts one day and calls the next is recognized as the same person
• Inbound webhooks are acknowledged within Meta's response window via API Gateway, then queued in Amazon SQS for asynchronous processing, decoupling ingestion from agent processing
• Voice calls route through Amazon Kinesis Video Streams for signaling and a managed TURN relay, requiring the only VPC-bound runtime in the stack (with a single NAT gateway for outbound traffic)
• Amazon DynamoDB stores customer profiles, orders, menu items, carts, and locations; Amazon Location Service handles geocoding and nearest-store lookups
• The entire stack deploys via AWS CDK, with agent containers built as ARM64 images through CodeBuild and stored in ECR

According to AWS, the design keeps the WhatsApp channel logic and backend ordering logic fully decoupled, meaning operators can add or remove channels (chat, voice note, voice call) without modifying backend code.

What This Means

This is not a new model launch — it's a systems integration pattern packaging existing AWS services (Bedrock AgentCore, Nova 2 Lite, Nova 2 Sonic, SQS, DynamoDB, Kinesis Video Streams) into a deployable reference for a specific vertical: restaurant ordering over WhatsApp.

The notable technical choice is using AgentCore's per-conversation microVM isolation combined with a single cross-channel memory store keyed by a hashed customer ID. That combination lets AWS claim session isolation for security purposes while still preserving continuity across channels — a common tension in multi-channel agent design.

The reliance on Meta's WhatsApp Business Platform (Cloud API, Media API, Calling API) as a prerequisite rather than something AWS deploys underscores that this is fundamentally an integration between two vendors' infrastructure, not a self-contained AWS product. Restaurant chains evaluating this pattern will still need a direct relationship with Meta's WhatsApp Business Platform, separate from their AWS account, before any of this architecture becomes operational.

For engineering teams, the more transferable takeaway is the pattern itself: decoupling channel ingestion (webhook plus queue) from agent execution (isolated runtimes) from business logic (a separate IAM-authorized API) is a reusable blueprint applicable well beyond food ordering, to any business wanting to expose an AI agent across multiple communication channels without rebuilding backend logic per channel.