aws-samples / sample-voice-ai-agent-studio

# Conversational AI Agent Studio
Design, build, and prototype voice AI agents on AWS. A guided studio for configuring real-time speech agents — from industry template to live phone call — powered by Amazon Nova Sonic, Strands Agents BidiAgent, and Amazon Bedrock AgentCore.
> **Sample / proof-of-concept.** This project is a reference implementation for learning and prototyping. Review and harden it (security, error handling, scaling, cost controls) before using it for production workloads.
## The Voice Agent Stack
A real-time voice agent isn't a single model — it's a stack of decisions. Each layer is chosen independently, and the trade-offs at one layer ripple into the others:
- **Model** — what hears, thinks, and speaks (speech-to-speech like Nova Sonic, or STT + LLM + TTS).
- **Framework** — what orchestrates the session: streaming, turn detection, and tool calls (Strands BidiAgent, LiveKit, Pipecat).
- **Integrations** — what the agent can actually do (knowledge bases/RAG, APIs, MCP, other agents).
- **Channel** — how callers reach it: the protocol and transport (WebSocket, WebRTC, PSTN, SIP).
- **Hosting** — where it runs and how it scales (managed runtimes, containers, on-prem).
For the concepts behind these layers, see Building Real-Time Voice Agents: Choosing Models, Frameworks, Protocols, and Hosting.
This studio turns those layers into a guided workflow: you pick a model and voice, wire up tools and RAG, choose a channel (browser, PSTN, or SIP), and deploy to a managed runtime on AWS — without hand-assembling the stack yourself. The current build uses Amazon Nova 2 Sonic (speech-to-speech) on the Strands BidiAgent framework, hosted on Amazon Bedrock AgentCore.
## Architecture
The solution has four deployable components. The core app is deployed by the CDK
(`deployment/deploy.sh`); the two telephony paths are deployed independently
because they provision internet-facing network infrastructure subject to a
higher security review bar.
| # | Component | How it's deployed |
|---|-----------|-------------------|
| 1 | **Frontend app** (React UI + demos/tools/RAG API, Cognito) | CDK — `deploy.sh` |
| 2 | **Voice agent on AgentCore** (Nova Sonic BidiAgent runtime) | CDK — `deploy.sh` |
| 3 | **PSTN relay** (Twilio TAC bridge) | Independent — `telephony/pstn/` |
| 4 | **SIP relay** (drachtio + bridge) | Independent — `telephony/sip/` |
## Features
### Agent Builder
- Guided wizard — choose template, pipeline, model, voice, tools, prompt, and workflow
- Industry templates — insurance, banking, contact center training, drive-through, automotive, healthcare
- Visual workflow designer with per-step tool assignment
- Live voice test with real-time transcript
### Speech & Reasoning
- Bidirectional streaming (speech-to-speech) — Amazon Nova 2 Sonic (no key needed), plus OpenAI Realtime and Google Gemini Live (bring your own API key)
### Integrations
- Built-in reserved tools (end call, transfer to human) available to every agent
- Custom tools defined in