Summary: A September 4, 2026 AWS post describes a WhatsApp ordering assistant built on Amazon Bedrock AgentCore and Amazon Nova 2 for text, voice notes, and calls. Separate channel runtimes call the same backend MCP tools, while AgentCore memory keeps continuity through a hashed customer identifier. The production-oriented design includes queues, IAM, and encryption, but AWS omits performance and cost benchmarks.
Ordering at a quick-service restaurant is spread across an app, a website, a phone line, and a counter — four doors, four separate systems to build and operate. Customer history scatters the same way, and the same person reads as an unknown on every channel. A September 4, 2026 post on the AWS Machine Learning Blog walks through an approach that consolidates ordering into WhatsApp on top of Amazon Bedrock AgentCore and Amazon Nova 2. What comes out is one assistant covering text, voice notes, and calls, with shared memory behind it, so a customer who texts one day and calls the next is recognized as the same person.
Architecture: A Single Backend Behind Three Channels
Three concerns are deliberately kept apart: the WhatsApp layer owns conversation, one agent runtime apiece serves each channel, and menu, carts, orders, and locations live in the backend. Inbound traffic lands on a single HTTPS webhook, is acknowledged with a 200 right away, and is then processed asynchronously through SQS — which is what lets every layer be deployed on its own.
The customer-facing door is the Meta WhatsApp Business Platform, which exposes the Cloud API webhook along with the Messages API, the Media API, and the Calling API. Everything on the AWS side, including API Gateway, Lambda, SQS, DynamoDB, and more, is provisioned by AWS CDK. Ordering logic reaches the agents as MCP tools published through AgentCore Gateway, a managed MCP server; agents invoke tools such as GetMenu and PlaceOrder by name, which keeps channel concerns clear of ordering logic.
Shared Memory Across Channels: What Makes Continuity Work
AgentCore memory supplies one shared, cross-channel record keyed by a hashed customer ID. When a message arrives, the worker derives a pseudonymous customer_id using a pepper held in Parameter Store. A runtime reads the customer’s long-term insights out of memory when a session opens, and writes events back when the session ends. An order can therefore begin in text and be completed on a voice call, with the agent fully aware of where the customer stopped.
How the Three Channels Flow: Text, Voice Notes, Calls
Each channel gets a runtime of its own while drawing on the same backend tools and the same memory. Text messages pass through the chat runtime, which streams Amazon Nova 2 Lite via the Converse API. Voice notes are decoded to 16 kHz PCM and handed to a bounded Amazon Nova 2 Sonic speech-to-speech session — with no transcription service anywhere in that path, it is true voice-in, voice-out. Voice calls ride WebRTC over a managed TURN relay from Amazon KVS, and the voice-call runtime is the only one that runs inside a VPC.
Points to Weigh in Practice
This is a substantial build. It draws on a long list of AWS services, and the pipeline compiles ARM64 agent images through CodeBuild. Standing up the Meta WhatsApp Business Platform is a prerequisite, and secrets have to be populated in Secrets Manager out-of-band. The design targets production, with dead-letter queues, IAM authorization, and encryption at rest all included.
For product builders, the takeaway that matters is keeping channels separate from backend logic. Publishing the backend as MCP tools means a channel can be added or removed without touching the ordering system. Shared memory, in turn, answers a genuine pain point: nobody wants to repeat themselves. Anyone exploring multimodal agents will find a concrete blueprint here for handling text, voice, and calls in one system.
One limitation: the AWS post omits performance benchmarks and cost estimates, so testing against your own traffic is the only way to find out. As a reference architecture, however, it is a solid starting point for a business that wants to meet customers where they already are.