I had parallel Claude agents write a strategy folder—ten documents, neatly cross-linked. The PR looked great until the automated reviewer caught 30 of 95 relative links pointing to nonexistent files. Claude conceded: "The automated review caught a real bug." The docs linked to invented filenames like `./market.md` instead of `market-analysis.md`. Skimming as a human, I would have approved it; every link looked plausible.

The problem isn't hallucinations alone—it's throughput. A single engineer now runs multiple agent sessions producing thousand-line diffs in minutes. Human review is the bottleneck. Two failure modes emerge: approve blindly or read everything and waste the speed gain.

Self-review fails because an author's bugs are plausible to the author. Research (Panickssery et al., NeurIPS 2024) shows LLM evaluators favor their own generations. Cross-provider review decorrelates errors: OpenAI's Codex consistently flags issues in Claude-written PRs that Claude misses—invented files, irrelevant code blocks, security violations. Codex misfires too, just differently. The power is multimodel: combining providers where models fill each other's gaps.

My pipeline: a GitHub Actions workflow (`.github/workflows/pr.yml`) triggers on PR open/push, collects diff and conversation, runs a pinned Codex call via composite action, posts findings as an updated comment, and publishes a commit status that stays red until resolved. The core is one LLM call wrapped in ordinary engineering—own the prompt, lifecycle, merge gating, and model pin.