AWS deploys a multimodal WhatsApp ordering assistant on Bedrock AgentCore
The reference architecture is a concrete demonstration of AgentCore's agentic capabilities applied to a real commerce use case. A single WhatsApp business number fields text messages, voice notes and live voice calls, with the agent maintaining a unified customer memory so a shopper is recognized whether they type or speak. Amazon Nova 2 provides the multimodal model layer, and the design deliberately decouples the messaging-channel logic from the ordering/business logic so each can evolve independently.
The shared-memory element ties into AWS's broader agent-memory push this week — the company simultaneously published lifecycle policies for AgentCore memory, explaining how to score, consolidate and prune agent memories on a nightly Step Functions workflow with a deployable CDK stack. Together they show AWS positioning AgentCore as a production framework, not a demo toy: persistent identity plus managed memory hygiene.
This is part of a dense week of Bedrock announcements — federated data-access patterns for agents via MCP, a native ServiceNow connector for Managed Knowledge Base, automatic sync scheduling, and Intuit's EWOK agentic disaster-recovery assistant. AgentCore payments also recently reached GA, letting agents autonomously pay for APIs via x402 and Coinbase. The competitive context is direct: Microsoft (Agent Framework + Cosmos memory) and Google (Managed Agents) are racing on the same agentic-infrastructure front, and AWS is differentiating on breadth of managed building blocks. The caveat for these blog-driven patterns is that they are reference implementations, not turnkey products; teams still assemble the pieces. Watch for AgentCore adoption metrics and whether Nova 2 multimodal quality holds against Gemini and GPT on voice tasks.
