Agent-to-agent pair programming

Source: Hacker News

Overview

What if you could let Claude and Codex work together as pair programmers, talking to each other directly? One acts as the main worker while the other serves as a reviewer.

Researchers at Cursor discovered that the most effective agentic workflows often resemble human collaboration. Their work on long‑running coding agents led to a multi‑agent workflow with a main orchestrator assigning tasks to workers—much like typical human teams. Claude’s “Agent teams” and Codex’s “Multi‑agent” features operate similarly, with sub‑agents reporting back to a primary agent. In the future, sub‑agents could even interact with each other, just as humans do.

While building a code‑review agent using Claude and Codex side‑by‑side, I noticed that they sometimes gave different feedback. Even when they agreed, their consensus provided a very strong signal. Our team addresses 100 % of the feedback when both reviewers agree. Code reviews are valuable because they occur on a multiplayer app where humans and agents collaborate, but they can slow down the feedback loop and become noisy.

Implementation

To make the feedback loop faster and more natural, I built loop, a dead‑simple CLI that launches claude and codex side‑by‑side in tmux, with a bridge that lets them talk to each other.

• The agents can be more proactive because their interaction feels natural.
• As the models improve, the collaboration should become even more effective.
• Since loop runs the interactive TUIs, you stay in the loop, can steer the conversation, answer questions, and follow up when needed.

Open Questions

• Should we split the work across multiple PRs?
• Should we share the PLAN.md in Git or in the PR description?
• Should we include a screenshot or video recording as proof of work?

Letting the agents loop can produce more changes than expected—usually a welcome side effect—but it can also make the human review harder.

Conclusion

Many people use multiple agent harnesses to avoid vendor lock‑in, contribute to open‑source projects, maximize subscriptions, or obtain diverse perspectives, strengths, and results. Multi‑agent harness applications should treat agent‑to‑agent communication as a first‑class feature. I’d love to see more tools adopt this approach.

Try it out:

Thanks to Léna Deloizy Delafosse, Will Horn, Tian Wang, and Ferruccio Balestreri for reading drafts of this.