Skills2026年4月9日·1 分钟阅读

OpenAI Realtime Agents — Voice AI Agent Patterns

Advanced agentic patterns for voice AI built on OpenAI Realtime API. Chat-supervisor and sequential handoff patterns with WebRTC streaming. MIT, 6,800+ stars.

OpenAI
OpenAI · Community
Agent 就绪

这个资产会安全暂存

这个资产会先安全暂存。复制的指令会要求 Agent 读取暂存文件，并在激活脚本、MCP 配置或全局配置前先确认。

Stage only · 29/100策略：需暂存
Agent 入口
任意 MCP/CLI Agent
类型
Skill
安装
Stage only
信任
信任等级：Community
入口
step-1.md
安全暂存命令
npx -y tokrepo@latest install 0d228731-33e3-11f1-9bc6-00163e2b0d79 --target codex

先暂存文件；激活前需要读取暂存 README 和安装计划。

TL;DR
Official OpenAI demo showing chat-supervisor and sequential handoff patterns for building production voice AI agents.
§01

What it is

OpenAI Realtime Agents is an official OpenAI demo showcasing advanced agentic patterns for voice AI. It demonstrates two key patterns: Chat-Supervisor (a realtime voice agent delegates complex tasks to a smarter text model like GPT-4.1) and Sequential Handoff (specialized agents transfer users between each other based on intent). Built with the OpenAI Agents SDK and WebRTC voice streaming.

It is designed for developers building voice-enabled AI applications, customer service bots, or multi-agent voice systems.

§02

How it saves time or tokens

The chat-supervisor pattern reduces token cost by using a fast, lightweight voice model for conversation flow while delegating expensive reasoning to a text model only when needed. This avoids routing every utterance through a large model. The sequential handoff pattern prevents prompt bloat by keeping each agent focused on a single domain.

§03

How to use

  1. Clone the repository: git clone https://github.com/openai/openai-realtime-agents.git
  2. Install dependencies: cd openai-realtime-agents && npm i
  3. Set your API key: export OPENAI_API_KEY=sk-your-key-here
  4. Run the demo: npm run dev and open http://localhost:3000
§04

Example

Chat-Supervisor Pattern:

User (voice) <-> [Realtime Voice Agent] <-> [Supervisor GPT-4.1]
                       |                          |
                  handles chat              handles complex
                  and simple tasks          tool calls and
                                            decision-making

Sequential Handoff Pattern:

User -> [Greeter Agent] -> [Sales Agent] -> [Support Agent]
             |                   |                |
        detects intent      handles sales     handles support
        and routes           queries           queries
§05

Related on TokRepo

§06

Common pitfalls

  • WebRTC requires HTTPS in production; localhost works for development but deployment needs proper TLS certificates
  • The realtime API has different pricing from the standard chat API; monitor usage carefully during development
  • Audio quality depends on network conditions; implement proper error handling for dropped connections

常见问题

What is the chat-supervisor pattern?+

The chat-supervisor pattern uses a lightweight realtime voice agent for conversation flow and delegates complex reasoning or tool calls to a more capable text model like GPT-4.1. This balances responsiveness with intelligence while controlling costs.

How does sequential handoff work?+

Specialized agents handle different domains. When a user's intent changes, the current agent hands off the conversation to the next appropriate agent. Each agent maintains focused context for its domain without carrying unnecessary conversation history.

What models are used in the demo?+

The demo uses OpenAI's realtime API models for voice interaction and GPT-4.1 as the supervisor text model. The Agents SDK orchestrates the handoffs and tool calls between models.

Can I deploy this to production?+

The demo is a reference implementation, not a production-ready service. You would need to add authentication, error handling, scaling infrastructure, and monitoring before deploying to production.

Does this work with non-OpenAI models?+

The demo is tightly coupled to OpenAI's Realtime API and Agents SDK. Adapting it to other providers would require replacing the WebRTC voice streaming layer and the agent orchestration SDK.

引用来源 (3)
TokRepo 相关
Multi-Agent FrameworksAI Agent ToolsFeatured Workflows
🙏

来源与感谢

Created by OpenAI. Licensed under MIT.

openai-realtime-agents — ⭐ 6,800+

讨论

登录后参与讨论。
还没有评论，来写第一条吧。

相关资产

OpenAI Agents SDK — Multi-Agent Workflows in Python

Official OpenAI framework for building multi-agent workflows. Handoffs between agents, tool calling, guardrails, tracing, and streaming. Lightweight, Python-native. 20K+ stars.

SkillsCLI Tools
OpenAI

OpenAI Agents JS — TypeScript Multi-Agent SDK

OpenAI Agents JS brings multi-agent workflows to TypeScript with provider-agnostic runs, zod schemas, tracing hooks, and sandbox agent patterns.

Skills
OpenAI

Groq Whisper — Sub-Second Speech-to-Text for Voice Agents

Whisper-large-v3 on Groq runs 166× realtime — 60-sec clip in <400ms. OpenAI-compat audio.transcriptions endpoint for voice agents.

Skills
Groq

OpenAI Agents SDK — Build Multi-Agent Systems in Python

Official OpenAI Python SDK for building multi-agent systems with handoffs, guardrails, and tracing. Agents delegate to specialists, enforce safety rules, and produce observable traces. 8,000+ stars.

SkillsCLI Tools
OpenAI
◈首页🔍搜索👤我的