Amazon Nova 2 Sonic enables real-time voice agent for healthcare appointments

Amazon has released a walkthrough for building a healthcare appointment voice agent using its Nova 2 Sonic speech-to-speech model on Amazon Bedrock AgentCore. Unlike traditional chained pipelines that lose vocal context through separate STT, LLM, and TTS steps, Nova 2 Sonic processes speech natively in a single model, preserving tone, hesitation, and urgency—critical in healthcare where a patient’s anxiety should change how the agent responds. The agent authenticates patients by voice, manages appointments (confirm, cancel, reschedule), collects pre-visit health information, and escalates to human staff when needed. It supports background noise, accented English, and multilingual conversations, switching languages mid-conversation without configuration. The implementation uses seven healthcare-specific tools built with the Strands Agents SDK, and runs serverless on Bedrock AgentCore with Amazon Cognito authentication, DynamoDB persistence, and SNS notifications.

The architecture includes a React frontend that streams bidirectional audio over SigV4-signed WebSocket connections to Bedrock AgentCore. For outbound calling, you integrate Amazon Connect or another telephony service. Amazon Nova 2 Sonic provides low-latency bidirectional streaming and tool-calling accuracy, enabling the agent to handle multi-step workflows (authenticate, query schedules, book) without introducing latency. The entire solution is designed to scale and reduce no-show rates—typically 5–30% in US healthcare—by automating reminder calls with natural conversation. The sample code and browser-based testing interface are available, making it practical for developers to adapt the agent to their clinic or hospital network's needs.