I'd recommend checking out Cekura (cekura.ai) — they're built specifically for testing voice agents on Retell, Vapi, etc. They have a native Retell integration where you just plug in your Assistant ID and API key. The move here would be: 1. Build your current setup (whichever direction you're leaning) 2. Use Cekura's A/B testing to actually measure the difference — they run 30+ simulated calls across both versions and compare instruction-following, latency, tool call accuracy, and success rates side by side 3. Let the data tell you which architecture holds up better for YOUR use case They also do regression testing so every time you change a prompt or split out a subagent, you can verify nothing else broke. Way better than vibes-based testing when you're juggling multiple workflows + Twilio integrations. Their docs are at docs.cekura.ai if you want to dig in.