Free to write, expensive to trust
Teams on agents write eight times the code and ship a third faster. The missing multiple is stuck at review, and clearing it is the job now.
The number pair that explains the year: teams running autonomous agents write roughly eight times more code than before, and ship about thirty percent faster. Thirty percent is real money. Set against eight times the output, it is a confession. The work is piling up somewhere, and it piles up in front of the merge button.
A bottleneck moves, it never leaves
Factory managers have known this forever: a line has one constraint, and improving anything else improves nothing. Typing got fast. Review did not. The queue simply moved.
Review cannot keep up, because eyeballs have a fixed throughput. Careful review tops out around four hundred lines of code an hour, a ceiling measured two decades ago and unmoved since. An agent hands you a ten-thousand-line diff on Friday. Reviewing it honestly takes a week. Nobody takes the week.
Machines cannot take it yet either. Put four AI reviewers on the same code and each one finds real bugs. Different bugs. No overlap. Each is worth running; none is worth trusting alone.
The question that splits the scores
Ask the strongest models “does this pass the tests?” and they score around eighty-seven percent. Ask what a maintainer actually faces, “would you merge this into your repo?”, and the score falls to the mid-fifties.
That gap is the whole story. The tests check what someone wrote down. Merging requires everything nobody wrote down: the style the codebase already carries, the side effects two systems over, the reason the old version did it the strange way. That knowledge lives in one head. The head is the bottleneck.
Write the judgment down
Anything cheap to verify eventually becomes training data. One review product turns every human thumbs-down into a written rule, and teams have accumulated tens of thousands of them. The moment judgment leaves a head and lands in a file, the agents start learning it, and the queue moves again. Operations and security are next in line, for the same reason: they are the best-documented parts of this job.
What resists being written down is knowing what a specific business means. “Keep track of my orders” is a requirement ten thousand products satisfy, and one of them fits a bakery. That meaning sits in no training set. It comes out in conversation, and it does not transfer. A calendar app and a scheduling app solve nearly the same problem and define good in incompatible ways.
What to do Monday
Review only what has blast radius. You cannot read everything, so stop pretending. Spend the scarce human attention on the code that can take the system down, leak the data, or move the money.
Turn judgment into artifacts. A spec, a rubric, an eval. If your best person’s sense of good exists only in their head, nothing scales and nothing survives their vacation.
Run several reviewers, never one. Four machines each catching different bugs is a useful panel. One machine is a coin with confidence.
Measure the gap. Eight times the code was never the point. Measure what shipped. Measure the distance between tests pass and maintainer would merge. That distance is the real backlog. And treat production as the last reviewer: grade what ships there, and feed every miss back into the rubric.
The typing is free. Trust is what’s left to manufacture, and none of it comes out of the model.
Working on something this applies to? Tell us about it — thirty minutes, no pitch.