Lesson 0001 · about 6 minutes
Why software factories fail
Not because agents cannot produce working code. Because the factory rewards what it can measure quickly—and maintainability is not one of those things.
The mechanism
Maintainability is a delayed system property.
A benchmark such as SWE-bench asks whether a patch resolves an issue without breaking known tests. That is useful evidence. It does not answer whether the patch preserves clear boundaries, keeps one concept in one place, or makes the next unknown requirement cheap. Those costs may surface months later as “shotgun surgery”: one conceptual change requiring edits in many places.
Harness engineering genuinely helps. Better tools, context, sandboxes, tests, and review loops raise the floor. But a harness cannot optimize a property for which it has no timely, reliable signal. More loops around an incomplete verifier optimize the incompleteness harder.
Retrieve before revealing
Scenario: An agent ships ten passing PRs per day. Three months later, every feature requires edits across six modules. Which capability was never verified?
Choose an answer.
The practical response
Keep humans at the high-leverage decision points: product intent, architecture, program shape, and review of small vertical slices. The goal is not to supervise every keystroke. It is to catch expensive design errors while the diff is still small.
Apply it now
Pick one agent workflow you use. Without reopening this lesson, answer:
- What fast signal declares success?
- What important quality can that signal miss?
- Where would you insert the cheapest human judgment?
Reply with your three answers. I’ll give immediate feedback and record the learning only after you demonstrate it.