How a Tiny Team Ships Like a Big One · The Product Track
A year in our GitHub stats
12% AI-generated code in October, 59% by January, 100% since April. What our own dashboard shows about the transition - and what the flattering numbers hide.
We keep a dashboard of our own engineering stats - PRs, commits, review flow, generated from the GitHub data and published internally as a static page. Preparing this post, I opened it and had to laugh: it was weeks stale. The dashboard tracking our automation was waiting for a human to remember to refresh it. (It runs on a schedule now. The fix for everything around here ends up being the same fix.)
Refreshed, it tells the story of our strangest year, and since this series has quoted its numbers piecemeal, the year-end post is the place to lay them out together - including the parts that flatter us less.
The lines that moved
The share of merged code written by AI: 12% last October, 59% in January, 100% since April. The steepest curve we've ever put on an internal chart, and an earlier post in the build track covers the operating model behind it. The dashboard adds what the summary number hides: the transition wasn't smooth. The curve has a plateau in late winter - weeks where trust stalled after bad experiences, before the review gates matured - and then a sharp elbow once the AI review bot shipped and the bottleneck moved.
PRs shipped: about 6x since January. The number everyone quotes back at us, so here's the honest anatomy of it. Some of that 6x is genuine throughput - more capabilities landing per week. Some is decomposition: agents ship smaller, more frequent PRs than humans batching a day's work, so "PRs shipped" partly measures a change in grain size. Both effects are real; the dashboard can't fully separate them; anyone who tells you their productivity multiple cleanly can is selling something.
Review latency: from hours-to-a-day to minutes. Last March, PRs sat waiting for the two people with context - overnight waits were normal, and a weekend PR met a Monday reviewer. That queue was our worst bottleneck and the AI reviewer deleted it: every PR now gets a full review in minutes, and only the risky fifth waits for a person. Of everything on the dashboard, this is the line that changed how work feels.
Commit velocity: recovered and steadier. Weekly commits dipped into the high double digits during our roughest patch and recovered past where they started once the rails and review gates landed - and the interesting part isn't the level, it's the variance. Machine-shaped work smooths the curve: agents don't have crunch weeks or vacation dips. The residual variance is the humans, which is as it should be.
What we actually watch
The vanity risk with all of this is real, so the numbers we manage by are mostly not the ones above: defects escaping to production (the fleet's zero-customer-reported weeks matter more than any throughput stat), time from issue to merged fix, revert rate on agent PRs, and human review turnaround on the changes that genuinely need humans. Throughput charts go up and to the right; those four tell you whether the throughput is good.
Real numbers
- 12% → 59% → 100%: AI-generated share of merged code, October to April - with a real plateau in the middle.
- ~6x PRs shipped since January - part throughput, part grain-size change, honestly inseparable.
- 80% of PRs reviewed only by AI; review latency from overnight to minutes.
- One stale dashboard, now refreshed on schedule like everything else we automate.
Where the humans sit
Reading the charts, not starring in them. The dashboard deliberately reports team-level aggregates - no per-person leaderboards, because instrumenting individuals is how you turn metrics into theater and teammates into gamers of throughput. The numbers exist to answer one question: is the system - humans, agents, gates - shipping more good software? Individual judgment calls stay in reviews, where they belong.
Steal this
Instrument your delivery before your AI transition if you can - our October baseline is what makes the whole story credible, including to us. Track PRs and commits, but pair every throughput metric with a quality check (escapes, reverts, review turnaround) so you can tell acceleration from recklessness. Keep it team-level. And schedule the refresh, because a stale dashboard is worse than none: it answers questions with the past.
On Thursday: the other dashboard - the one we pointed at our own meetings. It has a filler-word leaderboard. Nobody asked for it.
This post is part of the product track of How a Tiny Team Ships Like a Big One, a series on how six builders run a production AI company. Building at Aithon - if this is how you want to work, talk to us.