I thought I had verified it. Six days later I found out I hadn't — not all of it.
A post crossed my feed claiming that a very high share of engineers at one AI company run self-improving agent loops. I don't repeat numbers I haven't sourced, so before reacting I went looking for where it came from. That wasn't caution for its own sake: if the claim were true, I'd have had to rebuild how I run my own agent fleet.
I couldn't find that sentence in the public record.
What I actually found
Three separate, real things — none of which was the claim.
An engineer there is reported as saying he no longer writes code by hand, with a record day of about 150 pull requests. I only ever reached secondhand accounts of that, so I'm passing it along as a report, not as a measurement.
Anthropic's leadership has publicly estimated that 90% or more of their code is written by Claude, including scripts and experimental code. That's an estimate, and the source says so.
And there's a published measurement: as of May 2026, more than 80% of the code merged into their codebase was authored by Claude.
Every source described how much code the model writes. None described how many humans run the loop.
The viral version had swapped the subject — code became engineers — and borrowed "90%" from a different sentence about a different thing. Nobody had to lie for that to happen. A number just drifted one noun to the left.
Absence is the hard part
I want to be careful about what I'm claiming, because this is where fact-checks usually overreach.
I searched from several angles, pulled the top sources, and had each claim challenged by verifiers whose job was to refute it rather than confirm it. The verdict was unanimous. But unanimous about what? Not that the sentence doesn't exist. Only that I couldn't reach it.
I've been burned by this exact gap before. I once read a grep returning zero hits as proof that something wasn't there, and it was — my pattern was wrong. I've also had a search API's "no results" turn out to be a display cutoff rather than an empty set. Zero hits means your search stopped, not that the world is empty.
So the sentence I'll defend is "I couldn't find it in the public record," and not "it does not exist." Those are different claims, and only one of them is something I can actually support.
The search did hand me something useful, though. A published analysis of about 400,000 Claude Code sessions from roughly 235,000 people, between October 2025 and April 2026, found that people make about 70% of the planning decisions but only about 20% of the execution ones. I keep a human gate on every go/no-go call in my setup, and I'd half-assumed that made me slow to automate. Turns out it's roughly where the measured division of labour already sits.
Then my own note failed the same check
Six days later I was about to post a short version of all this. My rule is to re-fetch primary sources before anything with a number in it goes out, even when the research was already verified — days passing counts as drift.
Four claims in the draft. Three held. One didn't.
My research notes had recorded something like "only 0–20% of tasks can be fully delegated." That phrasing appears nowhere in the study. There's no "delegable" anywhere in it, and no 0–20% range. The most likely explanation is that I read "humans make about 20% of execution decisions" and quietly rewrote it as "0–20% of work can be delegated" — which is a different statement about a different thing.
Which is to say: the same swap I had just caught in someone else's number, in my own summary, six days later.
The check I built for other people's numbers never ran on my own.
What caught it wasn't the elaborate part. Not the multi-angle search, not the adversarial verifiers. It was one plain re-fetch of the source document, at the last possible moment, because a rule said to.
The gap wasn't verification. It was the return path.
Here's the part that actually changed how I work.
When I fixed that line, I fixed it in the outgoing post. The research note that produced the post stayed wrong. It sat there for six days, untouched since the day it was written, with no correction task filed against it — and I only noticed because I went back to write this article and checked.
My published work had a gate. The notes that fed it did not. Corrections travelled outward to the artifact and stopped, when the thing that needed fixing was upstream of everything the artifact would ever become.
That's fixed now, and I'd rather show it than assert it. The note carries a strikethrough on the wrong figure, a dated correction marker with the right one, and a line saying how it was caught — because silently editing a note leaves no evidence that anything was ever wrong. A second person re-fetched the source independently and confirmed the original text contains neither the word nor the range. And the missing step got written down as a rule: when you correct a number, immediately grep the whole workspace for the same claim and kill every copy.
Takeaways
For any number-shaped claim, confirm the subject and the denominator in a primary source before you let it move your roadmap. Most viral statistics aren't fabricated — they're real numbers describing something adjacent.
"I couldn't find it" and "it doesn't exist" are different sentences, and only one of them is verifiable. Say the one you can support.
Run your verification on your own notes, not just on other people's claims — and make sure the correction travels back to where the note lives, not just to what you published.
Where do your corrections go: into the post, or back into the note?
Sources: When AI builds itself (the >80% measurement and the 90%+ leadership estimate) · How Claude Code is used in practice (the ~400,000-session analysis and the 70%/20% split)
── Hideyuki Mori (Ayane International) 🔗 hideyuki-mori.com
Top comments (0)