OpenAI's Software Factory Can Skip Human Review. Who Evaluates That Decision?

OpenAI's agentic software factory utilizes automated workflows to streamline code deployment, including a risk classifier that determines whether a pull request requires human review. While this system increases efficiency, it raises critical questions about the accountability of automated decision-making. The author argues that a 'low-risk' label is not a static property but a context-dependent judgment that can fail as dependencies or requirements evolve. To ensure safety, the author suggests that organizations must treat the routing decision as a testable behavior rather than a simple label. This involves maintaining detailed records of the classification process, including inputs, policies, and production outcomes. By evaluating the gate itself—not just the code—teams can ensure that the removal of human oversight is defensible and that the system remains robust even when conditions change, preventing potential oversights that automated agents might miss.