agent-testing

Agent-to-Agent Voice Testing

Infrastructure for automated testing of Twilio voice applications using two AI agents that call each other.

Concept

Instead of manually testing voice apps by phone, use two WebSocket-based agents:

• Agent A (questioner/caller): Initiates conversation, navigates IVRs, asks questions
• Agent B (answerer/recipient): Responds to prompts, simulates customers or agents

Both connect via ConversationRelay WebSocket protocol and use an LLM (Claude, GPT, etc.) to generate natural responses based on configurable system prompts.

Architecture

┌─────────────┐    make_call     ┌──────────────┐
│ Test Script  │ ──────────────→ │ Twilio Voice  │
└─────────────┘                  │   Platform    │
                                 └──────┬───────┘
                          ┌─────────────┴──────────────┐
                     Parent Leg                    Child Leg
                   (outbound-api)                 (inbound)
                          │                            │
                    ┌─────┴─────┐              ┌──────┴──────┐
                    │ TwiML A   │              │  TwiML B    │
                    │ <Connect> │              │  <Connect>  │
                    │ <CR>      │              │  <CR>       │
                    └─────┬─────┘              └──────┬──────┘
                          │ wss://                     │ wss://
                    ┌─────┴─────┐              ┌──────┴──────┐
                    │  Agent A  │              │  Agent B    │
                    │ (port 8080)│              │ (port 8081) │
                    │ ngrok A   │              │  ngrok B    │
                    └───────────┘              └─────────────┘

Setup

1. Two WebSocket Servers

Each agent runs its own WebSocket server with a configurable system prompt:

const WebSocket = require('ws');
const Anthropic = require('@anthropic-ai/sdk');

const PORT = process.env.PORT || 8080;
const SYSTEM_PROMPT = process.env.SYSTEM_PROMPT || 'You are a helpful assistant.';

const anthropic = new Anthropic();
const wss = new WebSocket.Server({ port: PORT });

wss.on('connection', (ws) => {
  const messages = [];

  ws.on('message', async (data) => {
    const msg = JSON.parse(data);

    if (msg.type === 'prompt' && msg.last) {
      messages.push({ role: 'user', content: msg.voicePrompt });

      const stream = await anthropic.messages.stream({
        model: 'claude-sonnet-4-20250514',
        max_tokens: 256,
        system: SYSTEM_PROMPT,
        messages: messages,
      });

      let fullResponse = '';
      stream.on('text', (text) => {
        ws.send(JSON.stringify({ type: 'text', token: text }));
        fullResponse += text;
      });

      await stream.finalMessage();
      messages.push({ role: 'assistant', content: fullResponse });
    }
  });
});

console.log(`Agent listening on port ${PORT}`);