Simon Willison sets up a comparison: a traditional engineer typically ships 50–60 lines of production-ready, deployable code per day, around 200 lines on a great day. If a coding agent — an AI that can autonomously decompose and complete multi-step programming tasks — can ship 1,000 lines at equivalent quality, the efficiency gain is real, but the team's cognitive ceiling surfaces in parallel.
This is not a simple story that "more code means more efficiency." As coding costs fall, the human ability to understand code, audit coupling, and make architectural trade-offs becomes scarcer. That, we believe, is what deserves attention.
What this is
Willison is not dismissing line counts. He argues that when code is tested, maintainable, and shippable, lines still carry reference value; the moment those conditions drop away, lines only incentivize redundancy and feature stacking.
His notion of "conceptual integrity" — a system with clear boundaries, coordinated parts, and unsurprising behavior — therefore matters more. Coding agents can pile on features quickly, but they easily turn the codebase into something like the Winchester Mystery House with its 140 rooms: local capability multiplies while the cost of overall comprehension and maintenance climbs.
Industry view
What's compelling about this judgment is that it redefines productivity as delivery at equal quality, not as the speed of pressing the generate button. Skeptics counter that if companies turn line count into a performance metric, employees will manufacture longer code and more feature rooms.
Another risk: generation tools don't automatically erase debugging, security, dependency, and hidden tech-debt work. Willison's premise is parity in quality — yet test coverage, maintainability, and architectural consistency are exactly what get overlooked first and are the hardest to verify reliably.
Impact on regular people
For enterprise IT
Software output will rise, and IT departments' center of gravity is likely to shift toward review, testing, and architectural governance. The cheaper code generation becomes, the more visible the cost of unified standards and ship responsibility gets.
For individual careers
The speed advantage in writing boilerplate code will compress skill gaps; experience in framing problems, decomposing constraints, and reviewing AI output matters more. Junior developers may ship faster, but they won't necessarily build system judgment faster.
For consumer markets
App updates and small feature rollouts will likely accelerate, while fragmented design and maintenance debt may translate into lag, security issues, or higher subscription fees. Long-term stability will still determine whether consumers stay willing to pay.