March 26, 2026
When bots bicker, bugs whimper
Agent-to-Agent Pair Programming
Two chatbots argue over your code—and users love the chaos
TLDR: Axel Delafosse’s tool “loop” lets two chatbots, Claude and Codex, review each other’s code to speed fixes and signal agreement. Commenters split: some report dramatic bug reduction and automation wins, others demand proof, while one popular take casts Claude as creator and Codex as the professional complainer.
What happens when two AIs start bickering over your code? Developer Axel Delafosse dropped loop, a simple command-line tool that seats Claude and Codex side-by-side so they can talk, review, and fix like human pair programmers. The twist: when both bots agree on a change, Axel’s team treats it like a flashing strong signal.
But the real show is in the comments. One camp is all-in on bot duos: bradfox2 swears by 10–15 back-and-forth rounds and even wired it into GitHub so issues flow from error tracker to auto-fix pull requests (a pull request is just a code change request). Meanwhile, vessenes turned the thread into a meme with roles: Claude for creativity, Codex for bull‑headed complaining—like a good-cop, bad-cop routine for bugs. Others chime in with their own rigs, like Swival reviews and cook, saying even self-review with the same model can expose how rough first drafts are.
Not everyone’s buying the hype. cadamsdotcom loves the “vibes” but wants receipts: show the data, not just demos. The mood swings between “robots are my new coworker” and “show me the studies.” Either way, the peanut gallery’s having a field day with the idea that the fastest way to better code might be letting two chatbots argue while you sip coffee.
Key Points
- •Axel Delafosse introduces “loop,” a CLI that runs Claude and Codex side‑by‑side in tmux with a bridge for agent‑to‑agent communication.
- •The approach mirrors human collaboration patterns, building on Cursor’s multi‑agent orchestrator/worker research and existing features like Claude Code “Agent teams” and Codex “Multi‑agent.”
- •Dual reviews from Claude and Codex provide complementary feedback; consensus between them is treated as a strong signal that is fully addressed by the team.
- •Loop accelerates code review cycles, preserves context across iterations, and keeps the human in the loop for steering and Q&A.
- •Open questions remain about PR practices (splitting PRs, sharing PLAN.md, proof‑of‑work artifacts), and agent loops can increase change scope, complicating human review.