AI Drives 'Over-Engineering' as Software Build Costs Collapse… Factories Grow Faster Than Products

AI has collapsed the cost of building software, but this has resulted in over-engineering. As seen in Steve Yegge's case, AI agent systems mass-produce artifacts such as code, tests, documentation, and policies at near-zero cost, each seemingly valid in isolation. However, the cost of understanding, verifying, maintaining, and coordinating these artifacts remains high, creating a vicious cycle where generated outputs increase system complexity rather than delivering actual value.

Factories Growing Faster Than Products

Wheelhouse, an AI agent system Yegge is developing for his game 'Wyvern', includes approximately 50–60 agents and 18 executive roles. The system generates an average of 270 commits per day and prepared Wyvern for relaunch in less than 10 weeks. However, Wheelhouse's own codebase is approximately 600,000 lines, roughly twice the size of the Wyvern product itself (excluding game content). Yegge diagnosed that the factory is growing faster than the product, with the ratio heading toward 1:1.

The agents generated 450 legal artifacts and over 100 'fences' (rules), and even created a new role called 'Head of Wheelhouse Law' to manage governance. While this unexpectedly reveals a civilization-like structure emerging within the agent system, it also demonstrates the phenomenon of software factories directing their production capacity back onto themselves.

Production Accelerated, Value Stagnant

AI has removed the traditional friction required to produce artifacts such as policies, tests, and documentation. In the past, these efforts served as filters to screen out unnecessary generation, but AI bypasses them. As a result, token costs decrease, but cognitive load does not. The cost of identifying valid policies, detecting inconsistencies between documents, and understanding the causes of test failures remains high.

According to data from BetterUp, productivity increased after AI adoption, but the burden of review and revision also grew. Additionally, 40% of office workers reported receiving 'workslop'—plausible but practically valueless output. In Yegge's case, activity metrics such as agent count, token usage, and lines of code surged, but players requested a slowdown in the pace of change. This indicates that while production has accelerated, organizations and users are unable to absorb it.

The Vicious Cycle of Over-Engineering and Solutions

Dangerous over-engineering accumulates gradually. It follows a vicious cycle: agent misjudgment -> rule creation -> enforcement implementation -> exception occurrence -> exception documentation -> need for rule owners -> need for control over process change processes. Each step appears reasonable, but the entire system can become absurd. AI seeks completeness and tries to justify every addition, but has a weak tendency to keep the system clean for the sake of meaningless tidying.

The solution lies in recognizing that the volume of governance does not equate to quality. Every rule and artifact must address a named failure mode; if one cannot explain why outcomes would worsen if it were removed, it should be treated as sediment and eliminated. The most important metric is not the quantity of internal artifacts, but the actual improvements measured outside the factory.