Steve Gattuso, writing on his own site sjg.io, argues that AI has not simply made software cheaper to build. It has removed the friction that once forced teams to ask whether a policy, a test, or an agent role was worth creating at all. The scoreboard most teams are watching, commits shipped and lines produced, measures exactly the wrong thing, in his telling.
Gattuso’s case study is Steve Yegge’s Wheelhouse. It is a fleet of somewhere between 50 and 60 agents, 18 of them holding named officer titles, that Yegge pointed at his decades-old game Wyvern to get it back onto Android, iOS and Steam. In under ten weeks the fleet averaged 270 commits a day and rewrote the game’s production infrastructure. It also grew to roughly 600,000 lines of code and tests, nearly twice the size of Wyvern itself, and generated 450 internal legal artefacts: rulings, runbooks, and more than 100 rules Yegge calls fences.
The system eventually needed its own regulator. Yegge staffed a Head of Wheelhouse Law, whose job was retiring rulings that had gone stale and governing the governance. Gattuso treats this as confirmation of his thesis rather than a curiosity: a factory built to produce a product had started producing itself.
His shorthand for the underlying mechanism is that creation has become nearly free while ownership has not. A person still has to track which policy is the live one, catch the moment two documents start disagreeing, and work out six months on why a test began failing. Each fresh artefact becomes one more candidate for the truth. That cost does not show up in the commit count.
Gattuso backs the argument with outside data, though he is careful about its limits. Faros, a commercial engineering-analytics vendor, looked at telemetry covering upwards of 10,000 developers. Teams leaning hard on AI closed 21 percent more tasks and merged 98 percent more pull requests than their peers. Review time rose 91 percent, average pull request size climbed 154 percent, and bugs per developer rose 9 percent. Zoom out to the company level and Faros saw no meaningful link at all between how much AI was in use and whether results improved. That is correlational research from a vendor with a product to sell, not a controlled study, and Gattuso presents it as suggestive rather than proof.
A separate BetterUp survey, self-reported, put 40 percent of US desk workers as having been handed something in the previous month they took to be machine-written: finished on the surface, and quietly leaving the real thinking to them.
Stripped of the game-studio color, Gattuso’s argument describes a shift already visible across the industry. Cheap generation does not eliminate the cost of software. It moves the bill from writing the thing to maintaining, reviewing, and eventually deleting it, and that bill arrives on a longer clock than the one everyone is currently celebrating.
Yegge’s own numbers hint at where that bill lands. Even as it scaled, Wheelhouse piled up over 700 work items that were fully specified and never built, and Yegge reckoned the factory was eating 20 to 25 percent of every hour spent on Wyvern, a share he said showed no sign of shrinking. Gattuso calls this capacity seeking utilisation. Once a large agent fleet exists, keeping it busy becomes a goal on its own, independent of whether the product needs the output.
Gattuso is explicit that none of this argues against guardrails for agents operating without supervision. His objection is to mistaking the volume of governance for its quality, and to a fleet that grows its own coordination overhead faster than it grows the product it was built to serve.
For any team scaling an agent fleet past a handful of instances, this is a diligence checklist, not just an essay. Before adding another agent role, test suite, or policy layer, name the failure mode it addresses and who is responsible for retiring it later. Track hours spent reconciling AI output against a stated outcome, not lines of code or pull requests merged, before approving the next expansion of the fleet.
Steve Gattuso published this analysis, “AI Is Making Us Build Too Much,” on his own site, sjg.io, on September 3, 2026.
