AI Kills the Speed Limit — and the Software Engineering Middle Class

The core argument is that AI coding agents haven’t changed which engineering practices are good or bad — they’ve removed the natural rate limit on bad ones. Where a weak team once needed weeks to bury a codebase in unjustified abstractions, denormalized tables, and unnecessary infrastructure, an engineer can now generate 20,000 lines in an afternoon. The author paints a 2026 Monday of reviewing thousands of AI-authored diff lines, chasing a recurring bug nobody can trace, and discovering that a feature’s actual design decision lives buried in a rambling multi-round chat log rather than in anyone’s head.

The deeper problem is asymmetry: generating a bad decision is now nearly free, but reversing one is as costly as ever. Adding database tables takes ten minutes; migrating away from them once real users depend on the data means migration plans, failure handling, and orphaned-key risk — all while fresh PRs keep landing faster than anyone can understand them. Technical debt isn’t inherently bad, the author argues, but it only works when someone knows it’s a shortcut and can explain the tradeoff. That knowledge is exactly what’s evaporating when developers outsource their understanding to an LLM.

The economic conclusion is that AI widens the gap rather than flattening it. Companies pay six-figure salaries for judgment, not for turning specs into code — implementation was always the cheap part. Strong engineers become more valuable because AI multiplies their output without requiring a supporting cast; weak engineers become more expensive liabilities because they produce failure faster and can’t evaluate what the model hands them. The author expects pay to polarize toward the shrinking group who can actually be trusted to know what’s going on, and predicts the same squeeze will spread across knowledge work generally.