Why Does Beautiful Code Still Matter for AI Agents?

David Heinemeier Hansson still sweats the details when agents change the Ruby that runs the 37signals business. For 25 years the economic reason was simple. A coherent, malleable architecture let a small team change the software quickly, without an exorbitant cost and without a pile of bugs every time one edit touched another. Beautiful code is easier to understand, simpler, and more malleable. That case assumed humans would do the edits.

The payoff of that sweat is diminishing, and how far it still matters is an open question. For the moment it does, because tokens are scarce. Anyone without an endless budget is token limited. A system an agent can change without relearning the whole context keeps the next edit as cheap as the last one. The failure mode is the classic ball of mud, where nothing connects. He has seen agents produce it in his own code bases: the first pull request is mediocre, and five more stacked on top are not good. The argument is about this moment, not a forecast of what the tools will do in nine months or two years.