Technology

AI agents write more of the code, and review is slowing down

Two pieces of research from the past few weeks point at the same part of the work.

GitHub's Octoverse 2026 report says that by September 2026, more than 40% of pull requests on GitHub started with an agent's first draft, up from 17% in March. This is GitHub's own data about its platform. The report also says that "builders are showing up from every field".

A working paper by Fiona Chen and James Stratton, Artificial Intelligence in the Firm: Bottlenecks in Software Production, looks at what happens inside companies after they adopt AI agents. Its headline finding is that code review becomes the bottleneck. After adoption, the average time from submitting code for review to its approval rose by 56.1%, and the number of comments on each change rose by 38.6%. The paper has not been peer reviewed, and its data comes from Jellyfish, a company that sells engineering analytics.

What review looks like when you are not a developer

I am not a software engineer. I built my CMS and my CRM with Claude in Claude Code. In both, I wrote the requirements, made the decisions and tested the result. Claude wrote the code and the tests.

The paper studies teams of engineers reviewing each other's work, which is a different setting from mine. The question it raises is the same one I face: once code arrives quickly, someone has to establish that it does what was meant.

Thankfully I have had a fair bit of experience working with freelancers, and product/project teams to figure out what works and what does not to be able to pin point specific issues. I probably spend a LOT less time navigating these issues than most people in my position because of this. It still took a fair bit of my time, to be honest, but I was glad to put the time in because the output is exactly what I wanted.

My review rests on three things:

  • A written spec with settled decisions. The CMS spec had a table of decisions that were closed. It gave me something to point back to when a change quietly contradicted a decision.

  • A test for every change. Each CMS change comes with a new or updated test, the full suite runs, and then the production build runs. I described the routine in part 5 of the CMS series.

  • Seeing it run. Anything visible gets checked running: by Claude in a copy with sample data, then by me on the live system.

With the CMS, a simple thing like a CSS change was getting stuck in the change. All it needed was a code fix which I pointed out. In the CRM, certain fields were not working the way it should have because assumptions were made. I am of the opinion that assumptions are the death of all communications, and what kind of a communication coach would I be if I did not communicate these issues effectively. I did. The results speak for themself.

The limits of this kind of review

Checking behaviour catches a tool doing the wrong thing. It is weaker at catching a tool doing the right thing in an unsafe way, and security is where someone who does not read code is most exposed.

Owning the code also means owning every problem. As I wrote in part 5, when something breaks there is no vendor forum and no plugin update: there is me, the logs and the code. If you cannot check what an agent builds by any of these routes, or cannot give it that ongoing attention, a product someone else maintains is often the better choice.

When something was last built for you, by a person or a machine, how did you check that it did what you meant?

Teams using AI in their writing and documentation need review built into the work in the same way; AI Content Ops trains them to do it: aicontentops.online.

Send feedback

Goes to my inbox. Nothing is published.

All posts