Still Write Code in 2026? Yes — It's How Engineers Keep Their Attention and Taste

The author accepts the premise that software work has shifted from building software to building the “software factory” — the prompts, skills, tests, linters, type systems, and CI that let AI agents ship good-enough changes with minimal human oversight. He admits most of his own code is now AI-generated. But he pushes back on the conclusion that humans should stop writing code entirely and just review agent diffs from a distance.

His argument is about attention, ownership, and precision rather than raw coding ability. Reading and approving code, he finds, breeds passivity: slop slips through, fragility accumulates unnoticed, and micro-adjustments get harder — which ultimately hurts the agents too, since fragile code is as confusing to them as it is to humans. Writing code forces direct engagement with the execution environment instead of proxying every idea through under-specified English, letting an engineer think in precise, executable steps. Agents, he stresses, are not compilers; they’re more like freshly onboarded interns that read partial, sometimes-slopified code and generate conservative changes — meaning they tend to preserve and pile indirection around a human’s bad one-off decisions rather than deleting them (his example: state needlessly stashed in browser local storage that tripled the surrounding line count).

The takeaway is a hands-on middle ground. Humans should still spike approaches, clean up architecture, tighten weak testing strategies, and joyfully delete cruft — then let agents stamp out the resulting patterns. Just as a car factory occasionally has to tear down the assembly line or study an engine to find a 10% gain, engineers need the freedom to dive into low-level details while holding the big picture, rather than drawing arbitrary boundaries around what they’re allowed to touch.