Entering the Agentocene: A Statistical Investigation of AI-Era Coding | Just Another Unix User<br>Table of ContentsTL;DR<br>A Lesson from Epidemiology<br>A Short History of Not Writing Code<br>Naming the Epoch: The Agentocene<br>The Question: Is It Measurable?<br>Method: How I Analyzed 5,009 Repositories<br>ResultsFinding 1: AI Tools Went from Unmentioned to Unmissable<br>Finding 2: An Involuntary Signal Says the Same Thing<br>Finding 3: Does It Mean More Code? Not Provably, Not Yet

Implementation Aspects<br>Limitations<br>What This Means

TL;DR#<br>This started as a search for KPI differences from AI adoption. That led<br>to a more basic question first: can a shift like this be proven<br>statistically at all?<br>Software keeps handing more work to the machine: compilers, libraries,<br>autocomplete. AI is the next step, and this time it shows up in the data,<br>not just in opinions.<br>I mined 5,009 git repositories (Rust, Python, JavaScript, TypeScript, commit<br>history from 2019 through mid-2026) and tracked 10,812 individual<br>developers to look for a measurable trace of that shift.<br>It&rsquo;s there, from two signals that agree with each other independently.<br>Literal AI-tool mentions in commit messages went from essentially zero to<br>1 in 8 commits by July 2026. A much harder-to-fake signal, a sudden,<br>lasting jump in how long people&rsquo;s commit messages get, shows the same<br>accelerating curve for 15% of the developers with enough history to check.<br>Whether this is already producing more code, not just different-looking<br>commits, is a separate and harder question. My data doesn&rsquo;t answer it<br>cleanly yet, and I say so plainly further down.<br>This new era is already being called the Agentocene: agents do the work,<br>we no longer code by hand.<br>A Lesson from Epidemiology#<br>In June 2026, Europe was hit by a heat wave unlike any on record. The<br>question everyone asked: how many people died because of it?<br>That&rsquo;s where epidemiology runs into a familiar wall. For any single death,<br>it&rsquo;s nearly impossible to prove that heat was the cause. The death<br>certificate says heart failure, not 42 degrees.<br>The proof doesn&rsquo;t need single cases, though. Statistics answers the question<br>with excess mortality: you compare the observed number of deaths against the<br>number you&rsquo;d expect from previous years. When mortality climbs far above<br>that baseline, the heat has left its mark, no individual attribution<br>required.<br>EuroMOMO, the European mortality<br>monitoring network, publishes exactly this comparison every week: observed<br>deaths against an expected baseline and its normal range, pooled across 25<br>countries. The gap between the line and the band is the excess, no autopsy<br>needed for any single death in it.

I used the same principle to test for the Agentocene, the era in which AI<br>agents write our code. Look at a single commit and you can rarely prove an<br>AI produced it. Commits carry no reliable label, and AI-text detectors don&rsquo;t<br>work reliably on prose, let alone on code. But if AI agents really changed<br>how software gets written, that change has to surface as a structural<br>deviation in the aggregate data of many repositories, the same way a heat<br>wave surfaces in death counts, not in any one obituary. This post is about<br>finding that deviation.<br>A Short History of Not Writing Code#<br>Software has always been a story of programmers handing work to machines.<br>In the early decades, writing directly in assembly was the only option, and<br>every instruction, every register, was a manual decision made by a person.<br>Adding two numbers and printing the result took real work:<br>; x86 Linux syscalls: compute 5 + 7 and print the result<br>section .data<br>a dd 5<br>b dd 7<br>buffer db 10 dup(0)

section .text<br>global _start<br>_start:<br>mov eax, [a]<br>add eax, [b] ; eax = 12

mov ecx, buffer + 9 ; convert eax to ASCII, back to front<br>mov ebx, 10<br>.convert:<br>xor edx, edx<br>div ebx<br>add edx, '0'<br>mov [ecx], dl<br>dec ecx<br>test eax, eax<br>jnz .convert<br>inc ecx

mov edx, buffer + 10 - ecx<br>mov eax, 4 ; sys_write<br>mov ebx, 1 ; stdout<br>int 0x80

mov eax, 1 ; sys_exit<br>int 0x80

Twenty-some lines, and I still glossed over register allocation and the<br>calling convention. Here&rsquo;s the same thing after a few decades of delegation:<br>print(5 + 7)

Then came compilers, and plenty of programmers hated the idea. Trusting a<br>machine to decide which instructions to emit felt like giving up precision,<br>maybe giving up craft altogether.<br>It&rsquo;s worth remembering how big that jump actually was, and it happened to<br>come up again very recently. In May 2026, at the Open Source Summit in<br>Minneapolis, Linus Torvalds made almost exactly this point while pushing<br>back on claims that &ldquo;99% of our code is AI-written&rdquo; now circulating in<br>parts of the industry. As LWN reported,<br>his comparison was blunt: compilers increased programmer productivity by<br>something like a factor of a thousand. AI, in his estimate, adds maybe<br>another factor of ten on top of that. He wasn&rsquo;t dismissing AI: he pointed<br>out that Linux kernel submissions had climbed roughly...