# Agent Ping Pong: A Dual-Agent Collaborative Development Workflow Implemented via Clipboard Protocol

> Agent Ping Pong is an innovative dual-agent programming workflow that uses OpenClaw as the conductor and Codex as the executor, leveraging the clipboard as a communication protocol to enable AI collaborative development without direct API integration.

- 板块: [Openclaw Llm](https://www.zingnex.cn/en/forum/board/openclaw-llm)
- 发布时间: 2026-04-17T04:15:00.000Z
- 最近活动: 2026-04-17T04:20:19.642Z
- 热度: 159.9
- 关键词: OpenClaw, Codex, AI编程, 多智能体, 工作流, 剪贴板协议, 代码审查, 协作开发
- 页面链接: https://www.zingnex.cn/en/forum/thread/agent-ping-pong
- Canonical: https://www.zingnex.cn/forum/thread/agent-ping-pong
- Markdown 来源: floors_fallback

---

## Introduction: Core Overview of the Agent Ping Pong Dual-Agent Collaborative Development Workflow

Agent Ping Pong is an innovative dual-agent programming workflow that collaborates via OpenClaw (conductor) and Codex (executor), using the clipboard as a communication protocol to enable AI collaborative development without direct API integration. This workflow allows humans to act as relays to transfer structured information blocks, retaining the advantages of multi-agent systems while lowering technical barriers.

## Background: Needs and Existing Issues of Multi-Agent Collaboration

As AI programming assistants become more capable, a single agent can hardly meet complex development needs. The industry is exploring multi-agent collaboration models, but these usually require complex API integration or intermediary systems, increasing the barrier to use. Agent Ping Pong proposes a minimalist solution of "clipboard as protocol", using human relays to transfer information and solve the problem of integration complexity.

## Core Concepts and Role Division

The core concept of Agent Ping Pong can be summarized as "OpenClaw is the brain, Codex is the hands, and the clipboard is the communication protocol". Role division:
- OpenClaw (Conductor): Responsible for requirement understanding, technical solution formulation, code review, and quality control;
- Codex (Executor): Writes code according to specifications, implements functions, and fixes issues;
- Human User (Relay): Copies and pastes structured information blocks, acts as a communication bridge, and maintains supervision and final decision-making authority.

## Detailed Workflow

A typical development cycle includes five phases:
1. Requirement Clarification and Specification Definition: The user describes the requirement to OpenClaw, which generates a specification block;
2. Code Implementation: The user pastes the specification block to Codex, which implements the code and creates a PR (Pull Request) to return a report;
3. Code Review: The user pastes the PR report to OpenClaw, which generates a review block;
4. Iterative Fix: The user pastes the review block to Codex, which fixes the issues and returns an updated report;
5. Final Confirmation and Merge: Repeat until OpenClaw approves the merge, then the user instructs Codex to complete the merge.

## Technical Implementation and Toolchain

Dependent tools include:
- OpenClaw: An open-source AI assistant framework that supports local deployment and custom workflows;
- Codex Desktop: OpenAI's desktop programming assistant (available with ChatGPT Plus subscription), which can operate local files and integrate with GitHub;
- GitHub: For code version control and collaboration; Codex can automatically create branches, commit code, and initiate PRs;
- Vercel: An optional deployment platform that provides free hosting. The project repository includes a SKILL.md document with setup guides, workflow instructions, etc.

## Advantages and Applicable Scenarios

Advantages:
1. Zero integration cost: No need to configure API keys, webhooks, etc.;
2. Flexible and controllable: Humans can intervene to modify or terminate the process at any time;
3. Role specialization: Avoids context dilution in a single AI;
4. Progressive adoption: Does not change existing development tool workflows.
Applicable scenarios: Rapid prototype development, code refactoring, feature iteration, learning and exploration.

## Limitations and Considerations

Points to note:
- Context length limitation: Information blocks need to be within the AI's context window;
- Risk of information loss: Need to copy and paste completely to avoid omissions;
- Lack of real-time performance: Manual relay introduces delays, not suitable for scenarios requiring instant feedback;
- Tool dependency: Currently mainly adapted to OpenClaw and Codex; other AIs need to be explored independently.

## Conclusion and Insights

Agent Ping Pong is a creative open-source project that enables dual-agent collaborative development via the clipboard, worth trying for programmers. Its core concept "AIs write messages to each other, humans only relay" embodies a pragmatic AI application philosophy. This model demonstrates the possibilities of a new human-AI collaboration mode (humans coordinate multiple AIs), protocolized AI communication (extensible to other scenarios), and a decentralized AI ecosystem (indirect collaboration via humans), which may become a new norm in software development in the future.