Cross-Model Adversarial Agents

Two frontier models. Each one reviews the other's work.
Bidirectional adversarial review between Claude Code (Opus 4.8) and OpenAI Codex CLI (GPT-5.5) — every plan, every diff, every commit goes through the other model's blind-spot filter before it ships.

Quickstart · How it works · Usage guide · Product spec · Changelog

Maintained by VZT Tech Consulting · MIT licensed

---

The problem

When one model plans, implements, and reviews its own work, it does not challenge its own assumptions. It has systematic blind spots baked into its training data. A second model — trained on different data, with different priors — catches what the first one misses.

                       single-model loop                              cross-model loop
                       ───