Zen and the Art of AI-Assisted Development

Summary

Thane Thomson argues that discussions of AI-assisted development should focus more on quality, not only on software output or automation. Drawing on Robert Pirsig’s account of quality, the essay suggests that quality is not simply a subjective or objective property but a deeper force that guides decisions and change. Thomson says that, in practical use, LLM-driven coding agents frequently propose options that differ from his own judgment, even when using leading models. Hard pass/fail tests can make an agent’s output more reliable, and manual review offers a looser safeguard, but neither means that the agent itself can sense quality. The essay therefore treats humans as the quality-detection apparatus in software development and warns that “vibe coding” can produce superficially convincing but structurally poor systems. Thomson also questions whether recursive self-improvement toward intelligence or superintelligence is meaningful when intelligence and improvement lack clear definitions, linking intelligence to the ability to perceive quality. He argues that automation works best when output quality can be precisely measured, as in manufacturing, whereas software and prose require task-specific measures whose design still depends on human judgment. On that basis, he calls the idea of a fully automated software factory effectively dead when quality matters. Finally, the essay warns that easy delegation to coding agents could prevent junior developers from developing problem-solving wisdom and cause senior developers’ assessment skills to decay, making the preservation of human judgment a major challenge for the technology industry.