AI on Both Sides of the Pull Request

Abstract

Purpose: This article sets out an operating model for AI in software development with two separate duties. One AI instance creates the application change from specifications and requirements. A second AI instance counter-checks the resulting source code in the pull request against a fixed standards catalogue. The pace argument is direct: when humans both write the change and review the pull request, both of those tasks have to move to AI if the team is to keep the pace. A human stays in the loop on both sides, so the generator cannot open a pull request on unchecked intent and the two instances cannot revise each other without a bound. The article states how a later comparison of business outcomes would be structured across three scenarios: human-only development, AI generation with human review, and dual-AI generation and review with human-in-the-loop controls. Measured outcomes of that comparison lie outside this article. Method: The article is a design argument. It starts from the human baseline of an author and a separate reviewer, separates the two AI instances by duty, places a human checkpoint before the pull request and a human decision on the review, and states the stop rules that end a generate–review cycle. Measured delivery speed and defect rates lie outside this article. Contributions: (a) A two-instance model that maps specification and requirements onto generation, and the pull request onto an independent standards review. (b) A human-in-the-loop placement on both sides, with explicit stops so the cycle remains bounded. (c) A standards catalogue for the reviewer instance: specification conformance, Clean Code maintainability, ISO/IEC 25010 reliability, security, and maintainability as review questions, named security issues, and tests tied to the claimed requirements.