On a current project I’ve got Claude Code in the middle, two screens, with Gemini on the left and Codex on the right for their adversarial challenge.
The arrangement is deliberately uncomfortable for the model in the middle. Claude writes. The two on either side are there to argue — to question the design and hunt for whatever the author missed. They come from two other companies, with different training and different blind spots, and the blind spots don’t overlap. That non-overlap is the entire value of the setup.
The industry seems to be arriving at the same idea. This week OpenAI shipped an official Codex plugin for Claude Code, whose README says you can “use Codex from inside Claude Code for code reviews or to delegate tasks to Codex,” including a command called adversarial-review that “runs a steerable review that questions the chosen implementation and design.” OpenAI’s own framing is that it isn’t a separate runtime — “It is Codex, just invoked from inside Claude Code.” Gemini, for its part, has had an open-source agent in the terminal since last June, Apache-licensed, with a free tier of 60 requests a minute and 1,000 a day. The parts for a three-model desk are all cheap or free, and now they’re starting to talk to each other by design.
I thought I could dispense with the right screen. One critic seemed as though it ought to be enough. So — not as good.
