AI Is Making Us Build Too Much
Short summary
AI has made producing code, tests, policies, and organizational structures nearly free, but maintaining and understanding them hasn't gotten cheaper. This creates a trap where systems accumulate layers of infrastructure faster than they provide value, turning the factory into its own largest customer.
- Steve Yegge's AI agent system (Wheelhouse) grew to 600,000 lines of supporting code—nearly matching the 1.2M lines of the actual product itself—complete with constitutional governance, legal rulings, and a "Head of Wheelhouse Law" role, all created in under ten weeks.
- AI removes the friction that historically forced teams to justify new policies, tests, and documentation. What's cheap to generate becomes expensive to maintain: every new rule creates potential contradictions, every test needs monitoring, every document is a possible source of confusion.
- The real failure mode isn't obvious incompetence—each individual addition looks sensible. The problem is capacity-seeking-utilization: once you have agents producing work constantly, the system needs more coordination infrastructure to manage that work, which creates more things to maintain.
- Engineering telemetry shows the gap between activity and value: teams with high AI adoption completed 21% more tasks and merged 98% more PRs, but review time jumped 91%, PR size ballooned 154%, and bugs per developer rose 9%.
Full summary
AI has made producing code, tests, documentation, and organizational structures nearly free, but maintaining and understanding them hasn't gotten cheaper. The author uses Steve Yegge's account of Wheelhouse—an AI agent system managing his 30-year game project Wyvern—as the key example. In less than ten weeks, Wheelhouse's 50-60 agents generated 270 commits daily and helped push the game toward launch. But the factory itself has ballooned to 600,000 lines of code and tests, approaching the size of the actual product it manages. Yegge created 450 legal artifacts—offices, runbooks, fences, constitutional rules—and eventually had to hire a Head of Wheelhouse Law just to prune obsolete rulings. The system developed its own governance layer because stateless agents coordinating around the clock need explicit rules, authority boundaries, and institutional memory that humans carry informally.
The real problem isn't that Wheelhouse is obviously wasteful—it's producing legitimate work and may still be cheaper than the human infrastructure that would replace it. The danger is that AI removed the friction that used to force people to justify whether something was worth building at all. Policies, tests, documentation, reorganizations—all used to require effort, which acted as a filter. Now an AI can generate a complete policy, implementation, tests, procedures, RACI chart, and governance slide deck before lunch. What hasn't gotten cheap is ownership. Someone still needs to know which policy is current, notice when documents contradict each other, and debug a failing test months later. Each new artifact adds cognitive load and potential sources of confusion.
The over-engineering doesn't arrive as one stupid decision—it accrues through sensible steps. An incident triggers a rule. The rule needs a test. The test needs enforcement. The enforcement creates exceptions that need documentation. Eventually you're managing the management system. Yegge's agents grew the rulebook without curating it, elevating good craftsmanship into constitutional law. Humans over-engineered systems long before AI existed, but we had friction that constrained the impulse. Now that constraint is gone, and there's another dynamic at play: once you have a large fleet of agents, idle capacity feels wasteful even when idleness costs nothing. Work gets decomposed aggressively, backlogs swell, and the system demands more coordination just to absorb the constant change.