OpenAI's internal engineering teams have shifted a large share of routine coding work to Codex, their AI coding agent. Engineers describe handing Codex a task, walking away, and returning to review finished code rather than writing it themselves. The change is not marginal: senior engineers report spending more time reading and judging agent output than producing code. The shift exposes a design problem that most agent products have not solved. Reviewing agent-produced code is a different skill from writing code. Engineers need enough in view to judge well: what the agent was told, what it tried, what it changed, and why. Without that, review becomes slow or shallow. OpenAI's internal tooling appears to surface this information alongside the output, but the source does not detail how. The broader signal for product teams is about role change. The engineer who delegated the task is now also the reviewer of record. That means the hand-off from agent to human, and the review step itself, carry more weight than they did when engineers wrote the code themselves. Products that treat review as an afterthought, rather than the primary human moment, are building toward a gap.