Two AIs Beat One: How an OpenAI Plugin Lets Claude and Codex Argue With Each Other

OpenAI has released an open-source plugin that gives Claude Code a structured integration with Codex, OpenAI's own AI agent. Everything runs directly from VS Code through the Claude Code Extension, with no switching between interfaces.

A detailed hands-on write-up of this pairing has appeared on Habr. Drawing on his own experience, the author shows how to turn two AI advisers into genuine, structured opponents, and why this works even for tasks not directly related to code.

The problem this approach solves will be familiar to anyone who works with large language models. A single AI has no internal incentive to challenge its own conclusions. It is bound by its training conditions, prone to self-confirmation and, as a rule, does not back away from its initial position even when the arguments for it are weak. In the classic scenario, the user has to play the opponent: ask follow-up questions, request alternatives, and hunt for flaws in the proposed logic on their own. That takes time, expertise and constant critical attention, resources that are always in short supply in real work.

Two AIs from different vendors approach this differently. Claude was developed by Anthropic with an emphasis on safety and precise instruction following; Codex was built by OpenAI with a focus on deep understanding and generation of code. These models have different strengths, different blind spots and different principles for judging the quality of an answer. When one AI states a conclusion and the other challenges it from a fundamentally different training base, it is no staged performance but a clash of two different knowledge spaces. The result is often better than rephrasing the same query several times and putting it to a single model.

Before the plugin, managing a dialogue between two AIs was organizationally awkward: two open interfaces, manual copying of context between them, and constant checking that both models were working with the same data. OpenAI's new plugin removes this friction: Claude Code gets a direct channel to Codex, while the user stays in the familiar VS Code environment and can use both agents in a single workflow.

The author goes beyond simple integration and describes additional skills for Claude Code: custom instructions that set the structure of an AI debate. In his words, it is the skills that turn the interaction between the two agents from a chaotic exchange of remarks into a managed process with clear roles: one AI formulates and defends a position, the other methodically looks for weaknesses in it. Without properly configured skills, the first attempts to start a dialogue between the two models yielded little of value, which the author admits candidly.

There are plenty of practical scenarios: code review with cross-checking, generating competing architectural solutions, deliberate hunting for logical vulnerabilities in which one agent purposefully attacks the other's proposal. But the author's most interesting observation concerns non-technical tasks: when analyzing business decisions, weighing architectural trade-offs or developing complex arguments, two opposing AIs consistently produce a richer and less biased result than a single model, however strong it may be.

A multi-agent approach with different vendors is not just a doubling of compute but an attempt to build into the workflow something that used to come only from human expertise: an independent second opinion. Now it is available right in the IDE and, unlike a colleague, is never busy and always ready to argue.