This was an insightful read despite sounding like the usual shallow AI-pocalypse posting. I had a suspicion that the initial velocity that coding assistants made achievable would cool down rather quickly as we all are aware of the shortcomings of coding agents. The illustrated K confirms my gut feeling that the amount of steering, handholding and double checking would eventually consume more human ressources than a slower build process.

Perhaps current incentives and agent behavior will reshape in the near future - right now even the SOTA-models give me Golden Retriever vibes in their eagerness to please and work for me. The constant paranoia I developed when working in larger or complicated code bases has me apprehend “don’t code, talk to me” still, even after the advent of system prompts and MD-based skills.

Nice collection of sources and the author doesn’t gallop logically or starts evangelizing as the usual developer crowd tends to do when writing.