Historical source material recovered on 2026-09-09. Capture dates and source limitations apply; recovery is not fresh verification.

Voice Agents

AI agents whose primary interface is spoken conversation over a phone call, app, or in-browser audio stream. Earlier examples combine a transcriber, LLM brain, voice synthesiser, and tool layer; newer realtime audio models such as Gpt Realtime 2 collapse more of that loop into a native audio model that can reason, preserve context, and call tools.

Key points

Two top-level shapes: inbound (receptionist answering calls) and outbound (agent placing calls)
Front-end platform (Vapi) handles the voice stack; back-end workflow engine (n8n) handles deterministic business logic
System prompt is the single biggest lever on quality — expect dozens of iterations in production
Voice agents should self-identify as AI on the opening turn as a baseline ethical practice
Wireframing the conversational flow before building is strongly recommended — conditional logic branches quickly
OpenAI's GPT Realtime 2 demo shows a voice agent checking a calendar, staying silently in context until reactivated, and updating a CRM through tool calls ^[src-051].
For action-taking voice agents, Voice Agent Preambles help avoid dead air by acknowledging the task and updating the user while reasoning or tool calls run ^[src-051].
OpenAI's Build Hour reframes mature voice agents as Voice To Action Interfaces: the agent can operate an e-commerce UI or analytics dashboard through tools, not merely answer questions aloud ^[src-083].
Sierra's production notes add the enterprise reality: real voice agents need workflows, allowed tools, grounding, brand language, guardrails, VAD behavior, traces, redaction, payment-safe flows, and full-call simulations around the model ^[src-083].
Simmons's desktop-control walkthrough adds a local-computer variant: push-to-talk voice agents can operate browsers, MCP-connected applications, and accessibility-tree desktop controls, but they need explicit permission and cost boundaries ^[src-104].
^[src-202] adds three production lessons from the World's Fair Online Track: voice agents do not always need frontier models, the LLM should not blindly drive every part of the orchestration loop, and interruption handling is a first-class UX and reliability requirement.
The update reinforces that voice-agent quality comes from the harness around the model: state, routing, VAD behavior, deterministic tools, recovery paths, and evidence of what the agent actually did ^[src-202].

Related entities

Related concepts

Source references

^[src-007] Nate Herk cluster — Nate Herk — Voice AI agents cluster (4 videos)

– Videos referenced: zWLZ3bVVwD8, BO-jFbN4p8Y, y-cq_Qo4zVo, Qt3zMBH-FNg

^[src-051] OpenAI – "We’re introducing three audio models in the API" (2026-05-07)
^[src-083] OpenAI – "Build Hour: GPT-Realtime-2" (2026-05-13)
^[src-104] Pat Simmons – "GPT Realtime 2 Can Now Run Your Entire Computer (Just Your Voice)" (2026-06-17)
^[src-202] AI Engineer World's Fair Online Track 2026 playlist update (47 new transcript captures, 2026-06-22 to 2026-07-02)

Historical recovered snapshot — 2026-08-20

Preserved from ai-watch-src371-376-20260820-115500-correction. Documentary recovery on 2026-09-09; claims below retain their historical source scope and were not freshly verified. Current synthesis above remains authoritative.

^[src-012] Nate Herk cluster — Nate Herk — Voice AI agents cluster (4 videos)

– Videos referenced: zWLZ3bVVwD8, BO-jFbN4p8Y, y-cq_Qo4zVo, Qt3zMBH-FNg

^[src-051] OpenAI – "We’re introducing three audio models in the API" (2026-05-07)
^[src-083] OpenAI – "Build Hour: GPT-Realtime-2" (2026-05-13)
^[src-104] Pat Simmons – "GPT Realtime 2 Can Now Run Your Entire Computer (Just Your Voice)" (2026-06-17)
^[src-202] AI Engineer World's Fair Online Track 2026 playlist update (47 new transcript captures, 2026-06-22 to 2026-07-02)

Google Cloud’s “omni-app” pattern treats voice as a continuous perceive–reason–express loop spanning interruption, tools, browser action, memory and visual context ^[src-372].
Production quality depends on end-to-end interruption latency, deterministic tool boundaries, explicit state and degraded-mode recovery; continuous sensing also creates privacy, retention and cost trade-offs ^[src-372].

^[src-372] Annie Wang / Google Cloud Tech — “Real-time voice AI agents, explained (before you build one)” (2026-08-19)

Robin Cartier perspective

This page is part of Robin Cartier's working AI knowledge graph: a practical research layer for production AI, recommendation systems, experimentation, GEO, and agentic web readiness.

The useful next step is to connect this concept back to applied product leadership and operating models.

Recommended next

Keep reading from this thread

From 500 indexed pages and articles.