AWS has published a reference architecture for a restaurant ordering agent that handles WhatsApp text messages, voice notes and live calls through one business number. The three channels share the same ordering backend and customer memory, allowing a person who texts on one visit and calls later to retain continuity.
Meta’s WhatsApp Business Platform provides the customer-facing APIs. Amazon Bedrock AgentCore hosts three isolated agent runtimes, while Nova 2 Lite handles text and Nova 2 Sonic processes speech. The agents call menu, cart and order functions through a Model Context Protocol gateway. DynamoDB stores operational data, and AgentCore memory uses a hashed customer identifier across channels.
The design acknowledges incoming webhooks immediately and places work on an SQS queue, preventing longer agent processing from blocking Meta’s response window. Text replies pass through a sender function. Voice notes use direct speech-to-speech processing without a separate transcription service, while live calls relay encrypted media through Amazon Kinesis Video Streams.
AWS provides deployment code through its Cloud Development Kit, but the example is not a turnkey consumer service. Teams need a configured WhatsApp Business account, model access in a supported AWS region, permissions to create the required cloud resources and careful handling of customer identity, conversation history and payment-related tools.