Reference: Inbound Voice Support Agent
referencevoice-support
voice
Voice Infrastructure
framework
Agent Framework
memory
Memory
retrieval
Retrieval / Vector Store
eval
Eval
observability
Observability / Tracing
deployment
Deployment Target
mcp_servers
MCP Servers
ZE
zendesk
TW
twilio
Architectural notes
Reference architecture for a voice agent handling inbound support calls with handoff to human on escalation. Why these picks: - Haiku 4.5 as default for sub-second turn latency; Sonnet escalation only when the call requires deeper reasoning. - Vapi for end-to-end voice (STT + LLM + TTS + telephony) to avoid integration overhead. - Cloudflare Workers + Durable Objects for low-cold-start session state. Common gotchas: - Turn-taking quality matters more to user satisfaction than model smarts on most calls. - Mid-call model swaps add noticeable latency; design escalation as a handoff, not a swap.
Embed this stack
Show off your stack on your engineering blog or company page.
<iframe src="http://testserver/embed/stacks/voice-support-inbound" width="100%" height="380" frameborder="0" style="border: 1px solid #334155; border-radius: 8px;"></iframe>