Real-time voice agents with Stream Vision Agents and Amazon Nova 2 Sonic

This article explains how to build production-ready real-time voice agents by combining Stream's Vision Agents framework with Amazon Nova 2 Sonic and Amazon Bedrock.

• Amazon Nova 2 Sonic provides speech-to-speech AI with native turn detection and function calling capabilities
• Stream's Vision Agents abstracts infrastructure complexity with plugin-based architecture and global edge network
• Sub-500ms end-to-end latency achieved through distributed SFU media forwarding and optimized audio streaming
• Function calling enables agents to query databases, call APIs, and trigger workflows during conversations
• Complete working agent built in under 30 lines of Python code using decorators and async patterns
• Supports custom pipelines with separate STT/TTS providers like Deepgram and Cartesia
• Amazon Polly TTS integration available for non-realtime LLM architectures
• Ideal for no-screen environments (driving, field service) and high-volume phone support automation

Vision Agents simplifies building voice AI applications by handling connection management, audio encoding, and orchestration, allowing developers to focus on agent logic and user experience.