Dian

One github account has merged 7,797 public PRs since october 1st That's one every 3.4 minutes, day and night. AI agents broke PR counts, so i built a leaderboard where every number links to the github search behind it. look yourself up

Spent a lot of time this week pulling GitHub data for long posts. They got about 25 views each. The ones about my own Claude Code logs got 60. Noted.

github got 7.64M new accounts in september, about 3 a second. in january 2025 it was 2.58M, under 1 a second search makes it look like the new ones do less. only 13% of september's accounts have a public repo, against 30% for january 2025. most of that gap is age and a lagging index though. in random samples of 120 to 240 accounts per month, the share that made a public repo in its first 4 weeks was around 1 in 5 through 2025, 12 to 15% from april to august 2026, and 22% for september. search's repos filter missed 12 of the 35 september repo owners 200 random september IDs: - 43 already deleted or hidden - 30 machine or placeholder accounts - 7 orgs, 4 with a profile but no repo - 81 people with nothing public, no repo, no bio, default avatar - 35 with a repo, 10 of them a real project none of the 35 had an agent file or a lovable, v0 or bolt readme. anyone signing up to use copilot or codex on private code leaves no public trace, so this can't show how many came for the agents

Vercel's preview bot commented on 1.57M public PRs in September. In January 2025 it was 149k, so 10.5x. Cloudflare's grew 25x to 249k and passed Netlify in July. Netlify only grew 3.7x, while non-bot PRs grew 5.5x. Agent PRs lean Vercel. 12.5% of PRs with the Claude Code footer got a Vercel preview, vs 8.8% of all non-bot PRs. Cursor PRs were at 18.9%, Codex 10.6% and Copilot's agent 5.7%. About a third of all Vercel-previewed PRs had the Claude footer. It's not one big account either. I pulled 35 random Vercel PRs and they came from 35 different repos, 65% with 0 stars. About half of Cloudflare's comments now come from Workers Builds, which started putting per-branch preview links in PR comments in July 2025. Cloudflare's own bot also opened 19-31k PRs a month from March to July and previewed them. I left those out, same for AWS Amplify's test PRs. Vercel says agents went from under 3% of its deployments to about half by July. Only public repos, and only PRs a host's bot commented on. The footer and the bots can be turned off, so this is a floor.

in september the claude code footer was on 33% of non-bot public PRs in HTML repos, 30% in swift, 28% in typescript and 13% in java. in january those were 1.4%, 8.9%, 4.6% and 0.6% java being low isn't about claude. codex and cursor PRs sit near the bottom for java too. java PRs come from different places: in a random sample a third were in the author's own repo, against about half for typescript, and a quarter went to repos with 100+ stars like kafka and nifi. agent PRs mostly land in people's own repos half of january's HTML PRs were aws amplify test PRs titled "automated test pull request", which hid the jump. the HTML footer PRs are small edits to people's own sites and little web apps, a median of 79 lines across 2 files it's not a few heavy accounts. 60 random swift footer PRs came from 57 different people the footer can be turned off so this is a floor, language is the repo's main one, and it's a share of PRs, not of code

Public github PRs p/m went from 4.82M in january 2025 to 22.78M in september 2026. split into buckets, each PR counted once: - 5.9M carry an agent marker, the claude code footer or co-author line, or a claude/, codex/, cursor/ or copilot/ branch. a third of the growth, 26% of september - dependency bots like dependabot and renovate added 2.0M, fork sync and actions bots 1.1M - 9.0M, half the growth, have no marker a marker only shows an agent was involved, and the no marker half isn't mostly people either. in a random sample of 309, 16% had an agent in the commits (claude co-author trailers, cursor agent or jules as author) and another 36% read like agent writing, a judgment call on the text copilot went the other way. its cloud agent opened 317k PRs in march and 60k in september. github stopped it opening a PR unless asked on april 1 and moved copilot to usage based billing on june 1, and the drop lines up with both. claude code went from 1.0M to 4.4M over the same months

12.1% of public github PRs opened in september got a comment or review from an AI review bot. that's 2.77M out of 22.84M, up from 2.0% in january 2025. leave out PRs opened by bot accounts like dependabot and it's 14.3%, about 1 in 7 biggest by PRs touched in september: - coderabbit 1.11M - codex 898k - copilot review 621k - greptile 149k plenty of those aren't reviews though. in a random sample of about 50 PRs per bot from the last month, coderabbit flagged an issue on 11 and left only a notice on 31, mostly review skipped because a bot opened the PR or the repo has under 10 stars. codex flagged something on 18 and hit a usage limit on 14. copilot posted a quota message on 16 gemini code assist went from 201k PRs in july to 23k in september because google shut the free github app down on july 17, only the enterprise version still runs 4.15M of september's PRs (18%) have the "Generated with Claude Code" footer, so not opened by a bot doesn't mean a person wrote it

Any other dev dads out there with a family and a 9-5 trying to build some side income? How the fck do you do it? I'm overwhelmed trying to spend enough time with the kids and still work on a side project after work. Truly blessed my 9-5 is fully remote, but man, still.

my claude code sessions spend 56% of their time waiting on me. i split my local logs (308 sessions i started, most of them this month) into three buckets: the agent waiting for my next message, the model generating, and tools running. the model gets 31% and tools only 13%. gaps over 30 min count as me being away and are left out. if i cap them at 30 min instead, waiting goes up to 67% inside the tool time, bash is 53% and the chrome browser 15%. the median bash call takes 0.44 s, so it's thousands of short calls adding up. the longest single waits were a handful of shell commands that ran right up to the 10 minute bash limit my codex sessions split 47% waiting, 45% model, 8% tools. when codex is actually working, the model is 85% of that time. in claude code it's 71% this is only my own dev logs on one of my machines, not a study. the median gap before i send the next message is a bit over 2 minutes

The devs on my leaderboard merged 6,322 PRs on launch day. My site got 42 visitors. One of them was me. If you shipped something this month, what did your day 1 look like?

trustmrr ranks startups by revenue. i wanted that for developers so i built a monthly leaderboard of public merged PRs. look anyone up and see where they rank inspired by @marclou's trustmrr and @robj3d3's canivibecodeit mergemax.dev/?ref=x

mergemax.dev/?ref=x every count links to the github search it came from, so you can check any number yourself