The Voice-First Agent Orchestration Guide: Unifying GPT-Live, OpenAI Presence, and Realtime V3 in Codex CLI

Updated

codex-clivoiceGPT-LiveOpenAI PresenceRealtime V3agent-orchestrationenterprise

The Voice-First Agent Orchestration Guide: Unifying GPT-Live, OpenAI Presence, and Realtime V3 in Codex CLI

Three voice capabilities shipped within a fortnight of each other in July 2026: GPT-Live full-duplex voice on the ChatGPT desktop app1, OpenAI Presence as an enterprise voice-and-chat agent platform2, and streaming Realtime V3 conversations in Codex CLI v0.145.03. Each targets a different layer of the stack, but together they form a single voice-first architecture for agent orchestration — from a developer’s terminal to a production contact centre. This guide maps that architecture, shows how the pieces connect, and provides the configuration needed to wire them up.

The Three Layers

graph TD
    subgraph "Developer Workstation"
        A[GPT-Live Desktop Voice] -->|directs| B[Codex CLI TUI]
        B -->|Realtime V3 WebRTC| C[OpenAI Realtime API]
        B -->|spawns| D[Multi-Agent V2 Sub-Agents]
    end
    subgraph "Enterprise Production"
        E[OpenAI Presence] -->|GPT-Live telephony| F[Customer Voice/Chat]
        E -->|Codex improvement loop| G[Policy & Guardrail Updates]
        G -->|tested via| H[Simulations & Graders]
    end
    C -.->|same model family| E
    style A fill:#e8f4fd
    style E fill:#fde8e8

Layer 1 — GPT-Live: Full-Duplex Voice Control

GPT-Live-1, launched 8 July 2026, is a continuous audio model that listens and speaks simultaneously4. Unlike the previous Advanced Voice Mode’s turn-based architecture, GPT-Live processes audio continuously and decides many times per second whether to speak, listen, pause, or delegate to a background model5.

The hybrid architecture couples a lightweight, low-latency voice model (GPT-Live-1 and GPT-Live-1 mini) with GPT-5.5 running asynchronously for heavier reasoning4. When a query requires web search or deeper analysis, GPT-Live delegates to GPT-5.5 in the background whilst keeping the conversation flowing. The user experiences no pause.

On 23 July 2026, ChatGPT Voice — powered by GPT-Live — shipped on the desktop app for macOS and Windows1. For developers, this means voice-directed control over Codex tasks, Work threads, and Chat conversations from a single interface, with Appshots providing screen context on macOS6.

Layer 2 — Realtime V3: CLI-Native Voice Sessions

Codex CLI has supported voice since v0.105.0 (spacebar push-to-talk transcription)7, but v0.145.0 shipped streaming Realtime V3 conversations with audio inputs and audio tool outputs3. This is the bridge between GPT-Live on the desktop and the agentic loop in the terminal.

The realtime subsystem has evolved through three generations:

Version Transport Capability CLI Version
V1 WebSocket Quicksilver intent, basic voice v0.105.0+
V2 WebRTC Two-way audio, voice selection, progress streaming v0.119.0+
V3 WebRTC + Frameless Bidi Streaming audio I/O, tool audio outputs, handoff streaming v0.145.0+

V3 introduces frameless bidirectional support, meaning audio tool outputs (not just text) stream back during agent execution3. An agent can now speak its progress aloud whilst editing files — genuine pair-programming rather than dictation.

Layer 3 — OpenAI Presence: Enterprise Voice Agents

Presence, announced 22 July 2026, is OpenAI’s managed platform for deploying production voice and chat agents2. It brings together six components: policies and SOPs, guardrails, approved actions, simulations, graders, and a Codex-powered improvement loop8.

The improvement loop is the architectural link to the CLI world. Codex investigates signals from production sessions using a Presence plugin, identifies weaknesses, and proposes behavioural changes — but human staff must test and approve every change before rollout9. OpenAI reports 75% of inbound calls on 1-888-GPT-0090 resolve without human escalation, and the Codex review loop cut handoffs by 15% over a 10-day period8.

Configuration: Wiring Voice into Codex CLI

Basic Realtime V3 Setup

The [realtime] TOML configuration table, introduced in PRs #14556 and #146067, controls voice session behaviour:

[realtime]
version = "v3"
type    = "conversational"

[audio]
microphone = "MacBook Pro Microphone"
speaker    = "External Headphones"

[features]
voice_transcription  = true
prevent_idle_sleep   = true

Two modes are available7:

  • conversational — Interactive two-way voice with tool execution during audio turns. Sessions initialise with recent thread context.
  • transcription — Speech-to-text piped into the agentic loop as standard user messages. Lower latency, lower token cost.

Voice Selection and Device Configuration

The /audio slash command provides interactive device selection within the TUI10. Audio device preferences persist automatically in config.toml.

AGENTS.md Voice Directives

Voice sessions benefit from explicit behavioural guidance in your project’s AGENTS.md7:

## Voice Session Conventions

- Confirm destructive actions aloud before executing
- Summarise each completed file edit in one sentence
- If interrupted mid-sentence, restart the previous thought on next user turn
- Prefer short confirmations over long explanations when the task is clear
- When running tests, announce pass/fail counts rather than reading full output

These directives shape the agent’s audio responses during conversational mode, keeping voice interactions concise and actionable.

Voice-Directed Multi-Agent Orchestration

The real power emerges when GPT-Live desktop voice directs Codex’s Multi-Agent V2 system. A developer can speak a high-level objective, and the orchestrator spawns sub-agents to handle it:

sequenceDiagram
    participant Dev as Developer (Voice)
    participant GPT as GPT-Live Desktop
    participant Codex as Codex CLI
    participant Sub1 as Sub-Agent: Tests
    participant Sub2 as Sub-Agent: Refactor

    Dev->>GPT: "Refactor the auth module and update the tests"
    GPT->>Codex: Task created via desktop app
    Codex->>Sub1: spawn_agent(model="o4-mini", task="update auth tests")
    Codex->>Sub2: spawn_agent(model="o4-mini", task="refactor auth module")
    Sub1-->>Codex: Tests updated
    Sub2-->>Codex: Refactor complete
    Codex-->>GPT: "Both done. 14 tests pass, auth module refactored."
    GPT-->>Dev: Speaks summary aloud

With v0.145.0’s stabilised Multi-Agent V23, sub-agent model routing is configurable, preventing the cost explosion of uniform Sol spawning11:

[agents]
default_subagent_model    = "o4-mini"
hide_spawn_agent_metadata = false

[agents.profiles.heavy]
model            = "gpt-5.6-sol"
reasoning_effort = "high"

Voice + Goal Mode

Goal mode (/goal, shipped v0.128.0, default since v0.133.0)12 pairs naturally with voice. Speak a goal, walk away, and the Ralph loop pursues it autonomously — checking back via voice when budget thresholds are reached:

[goal]
token_budget        = 500000
continuation_prompt = "Continue pursuing the goal"
budget_limit_prompt = "Budget nearly exhausted. Summarise progress and stop."

Extending Voice: MCP and Hooks

Spokenly MCP for Mid-Workflow Voice Q&A

The Spokenly MCP server enables agents to request spoken answers during autonomous runs7:

codex mcp add spokenly --url http://localhost:51089

Add a directive to AGENTS.md:

ALWAYS ask clarification questions via the `ask_user_dictation` tool
from the spokenly MCP server, never as plain text.

This bridges voice input beyond initial prompts — the agent pauses, asks a question aloud, and waits for a spoken reply.

Text-to-Speech Output Hooks

Wire task completion to speech synthesis for full audio-loop closure7:

[hooks]
PostTaskComplete = "say -v Samantha \"$(cat | python3 -c \"import sys,json; print(json.load(sys.stdin).get('message',''))\")\""

On macOS this uses the built-in say command; on other platforms, substitute with espeak, piper, or a cloud TTS endpoint.

Enterprise Voice: From CLI Patterns to Presence

The architectural parallels between Codex CLI’s hook stack and Presence’s production governance are deliberate:

Codex CLI Concept Presence Equivalent Purpose
AGENTS.md Policies & SOPs Behavioural directives
PreToolUse / PostToolUse hooks Guardrails Runtime intervention
requirements.toml Approved Actions Permission boundaries
/goal + Ralph loop Codex Improvement Loop Iterative refinement
Named profiles Simulation scenarios Testing configurations
eval framework Graders Outcome verification

Teams prototyping voice agents with Codex CLI can expect their AGENTS.md conventions and hook patterns to translate structurally into Presence deployments, even though the runtime environments differ8.

Platform Constraints and Known Limitations

Voice support in Codex CLI carries several constraints worth noting710:

  • Linux: No native voice transcription. Use Spokenly MCP or WhisperTyping as workarounds.
  • Windows CLI: Realtime voice requires the Desktop app; the CLI TUI alone is insufficient.
  • Session length: Long voice sessions trigger context compaction. Ensure compact_prompt preserves audio transcripts.
  • Clip duration: Spacebar dictation caps at 60 seconds per clip.
  • Echo loops: If transcriptions feed back into prompts, rapid token consumption follows. The backpressure and silence detection fixes in v0.116.0–v0.117.0 mitigate this but do not eliminate it entirely.
  • Presence availability: Limited general availability only, with deployments led by OpenAI Forward Deployed Engineers. No self-service tier exists yet2.

What Comes Next

The convergence of GPT-Live, Realtime V3, and Presence points toward a future where voice is the primary orchestration interface for coding agents. The Codex Micro hardware control surface ($230, six illuminated keys) already ships with a dedicated voice-activation key13. As Realtime V3 matures and Presence opens beyond its current FDE-led deployment model, expect voice-directed agent orchestration to move from novelty to default workflow.

For now, the practical path is clear: configure [realtime] for V3 conversational mode, add voice directives to AGENTS.md, wire up Spokenly for mid-task Q&A, and start talking to your agents.

Citations

  1. OpenAI, “ChatGPT Desktop 26.715 Changelog — ChatGPT Voice powered by GPT-Live,” 23 July 2026. https://learn.chatgpt.com/docs/changelog ↩ ↩2

  2. OpenAI, “Introducing OpenAI Presence,” 22 July 2026. https://openai.com/index/introducing-openai-presence/ ↩ ↩2 ↩3

  3. OpenAI, “Codex CLI 0.145.0 Release — Audio inputs, Realtime V3 streaming, Multi-Agent V2 stabilisation,” 21 July 2026. https://learn.chatgpt.com/docs/changelog ↩ ↩2 ↩3 ↩4

  4. MLQ News, “OpenAI Launches GPT-Live-1, a Full-Duplex Voice Model That Listens and Speaks Simultaneously,” July 2026. https://mlq.ai/news/openai-launches-gpt-live-1-a-full-duplex-voice-model-that-listens-and-speaks-simultaneously/ ↩ ↩2

  5. BuildFastWithAI, “GPT-Live Review: OpenAI’s Full-Duplex Voice Model Explained,” July 2026. https://www.buildfastwithai.com/blogs/gpt-live-review-openai-voice-model-july-2026 ↩

  6. VentureBeat, “Agentic coding goes hands-free as OpenAI brings GPT-Live’s full duplex voice control to Codex and ChatGPT on the desktop,” July 2026. https://venturebeat.com/orchestration/agentic-coding-goes-hands-free-as-openai-brings-gpt-lives-full-duplex-voice-control-to-codex-and-chatgpt-on-the-desktop ↩

  7. Daniel Vaughan, “Codex CLI Realtime Sessions: Voice Pair Programming, Transcription Mode, and the [realtime] Config,” 31 March 2026. https://codex.danielvaughan.com/2026/03/31/codex-cli-realtime-sessions-voice-transcription/ ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7

  8. Help Net Security, “OpenAI Presence connects AI agents to enterprise data with built-in guardrails,” 22 July 2026. https://www.helpnetsecurity.com/2026/07/22/openai-presence-ai-agent-platform/ ↩ ↩2 ↩3

  9. OpenAI Developers, “Build an Agent Improvement Loop with Traces, Evals, and Codex.” https://developers.openai.com/cookbook/examples/agents_sdk/agent_improvement_loop ↩

  10. Daniel Vaughan, “Voice-Driven Development in Codex CLI: From Push-to-Talk to Realtime V2 WebRTC,” 17 April 2026. https://codex.danielvaughan.com/2026/04/17/codex-cli-voice-realtime-webrtc-push-to-talk/ ↩ ↩2

  11. Daniel Vaughan, “Sub-Agent Model Routing in Multi-Agent V2,” 24 July 2026. https://codex.danielvaughan.com/articles/2026-07-24-codex-cli-sub-agent-model-routing-multi-agent-v2-hidden-schema-cost-control-configuration/ ↩

  12. Daniel Vaughan, “Goal Mode: How Codex CLI Turns a Single Objective into Hours of Autonomous Work,” 23 July 2026. https://codex.danielvaughan.com/articles/2026-07-23-codex-cli-goal-mode-long-horizon-autonomous-workflows-ralph-loop-token-budgets/ ↩

  13. Rohit AI, “OpenAI Codex Micro Puts the Agent Control Loop on Your Desk.” https://rohitai.com/blog/openai-codex-micro-agent-control-surface ↩