---
title: "Noted on 8 Oct 2026"
author: Marc Roberts
date: 2026-10-08
url: https://www.marcroberts.info/notes/2026-10-08-who-read-it/
---

# Noted on 8 Oct 2026

For as long as I've been reading other people's codebases, my first question has been who wrote this, and do they still work here. It feels like that's changing, and a more useful question now is who read it.

When a coding agent writes most of the code, authorship tells you little, so the approval is what's left to look at. An approval is a claim that a person understood a change well enough to be answerable for it, and unlike many claims, you can test it from the repository: how big the pull requests are, how long they waited, whether the approver ever said anything, and how many merged with nobody looking. Like onboarding time, it's already happened, and nobody can tidy it up after the fact.

There's now a careful study of what this looks like from the inside. Researchers at Carnegie Mellon and Stanford followed one AI-forward software company (802 developers, 196,000 pull requests) through a mandate to double its engineering output. Output per engineer did double, and about 90% of pull requests ended up AI-authored. But pull request volume tripled while the pool of reviewers grew by half, so each reviewer's load doubled. The share of pull requests that any human reviewed fell from 89% to 68%. The share where a human wrote so much as one comment fell from about 39% to 21%. What grew was the silent approval, which roughly doubled per reviewer. Automated review went from about 19% of pull requests to 84%.

Nothing visibly broke either. Merge rates held, and reverts went down if anything. The authors call this "evidence against an acute quality collapse, not a clean bill of health", which is carefully hedged and, I think, spot on.

Investors seem to agree on where the bottleneck is. In August, CodeRabbit raised $143m at a $1.5bn valuation, co-led by Atomico, to be "the control layer" for changes written by humans and agents alike. So the answer to more machine-written code than anyone can read.. is a machine to read it. I'd want one too, and on the evidence above it seems to work.

The risk for an investor is different, though. What you're buying is a team that can change the system on purpose, and that depends on somebody understanding it. A silent approval ships the change and teaches nobody anything. You find out the day something odd happens in billing and there's nobody to ask, because nobody ever really knew.

These days the first thing I look for is which parts of the system a human has demonstrably read, and which parts the company is taking on trust.

If your team ships mostly agent-written code, what does an approval mean at your place now? Is it documented somewhere?