Restraint Becomes the Quality Gate When AI Makes Features Cheap

articleaisoftware-qualityproduct-designagentic-software

Turns joelclaw's agent-generated feature work into concrete gates: reliability first, narrow purpose, low attention cost, and fast repair.

AI didn’t invent bad software. It increased the speed and volume at which we can make it. In “Quality Software”, Alasdair Monk argues that the old quality bar still applies even when implementation gets cheap.

Monk’s definition starts below taste and polish: quality software works, stays quiet, knows its limits, and gets fixed fast. “Restraint is respect” is the sharpest part. Every button, feature, notification, and megabyte makes a claim on the user’s machine or attention. Cheap feature production makes saying no more important, not less.

The article also draws a hard boundary around AI-generated writing. Agents can build software, Monk says, but words intended for a person should remain human-written. Whether that boundary holds everywhere is debatable, but the framing is useful: sloppy interface copy is not only a quality failure; it is a failure to respect the person reading it.

For joelclaw, this gives agent-generated work a better test than “does it ship?” A feature should have a narrow purpose, avoid demanding attention, survive ordinary failure, and make repair cheap. More generated code without those gates is only faster slop.

Key Ideas

  • AI increases the production rate of weak software; it does not create the underlying quality problem.
  • Reliability comes before polish. A handsome interface that breaks still fails the quality test.
  • Quiet software does its job without badges, banners, tours, or other demands for attention.
  • “Restraint is respect” turns scope control into a user-facing quality rule rather than an internal project-management preference.
  • Good software has explicit limits of size and purpose instead of racing to become “agentic” everywhere.
  • Fast repair matters because no system is unbreakable; quality includes what happens after failure.
  • Monk treats human-facing words as a boundary for automation: agents can build, but people should write for people.
  • In a reply, Adam Munns applies the same restraint to generative UI: direction and taste should come before producing 100 variants.
  • Alasdair Monk cites SerenityOS, created by Andreas Kling, as a recurring design reference.