As AI coding agents accelerate code generation, the bottleneck shifts to reviewing the volume of AI-produced code. The author often rejects AI-generated code even if it works, citing the need for scalable and maintainable solutions.
With AI coding agents accelerating code implementation, the cognitive burden of reviewing AI-generated code has become a new bottleneck. The author often rejects AI-proposed code even if it passes tests and CI, citing inadequacy in scalability or maintainability.
In the past, developers would explore the codebase, experiment with solutions, and implement themselves, leading to high confidence when submitting PRs. With AI agents, completing large tasks still takes days, but the author frequently rejects all changes from the first session and starts over. The improvement in the second session comes not from a better model but from a person who has consolidated the problem and can guide the agent better.
Accepting AI-generated code just because it 'works' is risky. Engineering is about implementing adequate, scalable, and maintainable solutions. Current AI agents cannot autonomously perform tasks in a sustainable manner without guidance from a skilled engineer. Therefore, human review alongside AI review is essential.
Comments sympathize with the author's stance of rejecting AI code they cannot understand or control, but also note that for low-risk internal tools or static websites, using AI code might be acceptable. A strong opinion emerged that if an engineer on call says 'the AI wrote it' when troubleshooting, they should be fired. The core debate revolves around maintainability and accountability rather than whether the code works.