Imagine if two AI tools, Claude and Codex, could work together like human programmers. This idea is becoming a reality with new developments in AI workflows. Researchers at Cursor have created a system where these two AIs can collaborate directly. They act as pair programmers, with one as the main worker and the other as a reviewer.

This approach mimics how human teams operate. In traditional settings, team members often share tasks and provide feedback to each other. Similarly, Claude Code “Agent teams” and Codex “Multi-agent” features allow subagents to report back to a main agent. This structure opens up possibilities for more natural interactions between the AIs.

Improving Feedback Loops

While experimenting with this setup, I discovered something intriguing. When Claude and Codex reviewed code together, they provided different feedback. Even when their suggestions aligned, it was beneficial rather than redundant. Our team found that we addressed all feedback when both reviewers agreed.

However, traditional code reviews can slow down the process. They often become noisy with too many opinions. To tackle this issue, I developed `loop`, a simple command-line interface (CLI) that runs Claude and Codex side-by-side in tmux. This tool allows them to communicate directly, speeding up the feedback loop while maintaining context across iterations.

The Future of Agentic Workflows

The future of these workflows may resemble familiar teamwork rather than automated processes. As we refine this method, several questions arise about enhancing human involvem

ent in code reviews. For instance, should we split tasks across multiple pull requests? Or should we include a PLAN.md file in our Git repository?

Additionally, sharing visual proof of work—like screenshots or video recordings—could improve clarity during reviews. While letting agents collaborate can lead to unexpected changes that are often welcome, it also complicates human review processes.

Why Multi-Agent Collaboration Matters

Many users are exploring multiple agent harnesses for various reasons. Some want to avoid vendor lock-in or contribute to open-source projects. Others seek diverse perspectives from different AI tools like Claude and Codex.

As multi-agent applications evolve, treating agent-to-agent communication as a key feature will be essential. This shift could significantly enhance user experience by making collaboration smoother and more effective.

Key takeaways

• Claude and Codex can work together like human programmers.
• The `loop` CLI speeds up feedback loops between AIs.
• Future workflows may resemble traditional teamwork more closely.
• Sharing visual proof of work can clarify code reviews.
• Diverse agent harnesses offer unique perspectives on coding tasks.

FAQ

• What is `loop`? It’s a CLI tool that connects Claude and Codex for faster feedback.
• How do Claude and Codex interact? They communicate directly to provide collaborative code reviews.
• Why use multiple agent harnesses? To gain different insights and avoid vendor lock-in.

For the original report, see the source article.

Sources

• Original source