GPT (Codex)

The model the second executor runs on. It handles the heavy autonomous coding — sweeping refactors, multi-file features, grind-to-completion build loops — and plays devil's advocate on completion claims, attacking the claim with fresh eyes and no stake in it being true.

A different lab's model has different blind spots than the ones that write my code. That independence is what makes its review pass worth having — it fails differently, so it catches differently.

It gets a written spec, a bounded scope, and room to grind. Its output never ships unreviewed: the primary session reads the diff, runs the tests, and looks at the result. Two strong models with different failure modes catch more than either alone.

Comes with the Codex CLI (npm i -g @openai/codex). The transferable idea is a second frontier model as executor and skeptic, not the specific one.