If you’re still shipping like you used to, or just seeing a percentage increase from where you used to be, you’re not doing anything interesting yet. A step change is the signal. This is the story of ours.
There are two different things you can mean when you say AI in engineering. “I build software that uses AI” is a product decision. “I use AI to build software” is an operating decision. A senior candidate recently failed our interview on exactly that distinction. Valuable. Not this.
Valuable. Not this.
This is the one this story is about.
I wrote that distinction down on July 24, 2025, before any of our current job postings went up. It hasn’t changed.
From August through December, our team excluding me averaged 3.3, 9.2, 11.5, 7.5, and 8.5 merged PRs per developer per week. Under twelve the entire time. The machines were being built, but features weren’t shipping faster.
August through December 2025, under twelve. June 2026, 62.2. Bars scaled to a 75-PR floor.
If you’re five weeks in and nothing has moved, you’re looking at the part of our chart that looked like failure. That part lasted five months.
On October 16, 2025, Anthropic shipped Agent Skills. I found their plugin system the same day, and the first commit landed in our own repo five days later. That’s where the tooling started.
November was the inflection point. My numbers went from 21 to 62 to 81, but the team didn’t move with me right away. Excluding me, they went from 11.5 to 7.5 to 8.5. They started moving in January — 24.7, 29.3, 32.9, then 45.1. By June, they were at 62.2.
The tool worked for the person who built it first.
The team didn’t move with me right away.
They started moving in January.
By June, they were at 62.2.
Capability adoption lagged capability arrival by about two months. The tool worked for the person who built it before it worked for everyone else. That’s the most useful thing I can tell you if you’re trying to copy this.
But the numbers only tell you that something changed. The more useful question is what we actually built.
You draft the work once. The line builds it and ships it. That’s the machine at v1. We’re running v0.8.
Define done. Decompose the work. Write the acceptance criteria. AI doesn’t touch this room because this is the constraint on everything downstream.
The automation takes the drafted spec and builds it. This is the part people usually imagine when they think about AI coding. For us, it’s logistics.
Review, verification, and merge are mostly automated now, with a rework loop back to the drafting table when something comes off the line wrong.
Draw those as three equal boxes and you’ve already misunderstood it. The drafting is the constraint. I don’t care if it’s a two-line change. I care that you follow the process because you want as much determinism in the non-determinism as you can get.
The pattern was simple: get a skill working, break it into parts, automate those parts, then hand the whole thing to an agent.
Ticket to pull request landed November 14. Pull request to merged landed November 27. Both were automated in the same week in January. They started thirteen days apart and reached the automation step one day apart. This was one program moving across the whole pipeline, not a pattern we noticed later and worked backward to explain.
Automated the same week in January.
Thirteen days apart at the start. One day apart on the automation step.
Each rung also leveled up the person who built the next one. Skills didn’t make the team faster in October. They made me faster, which gave me the capacity to build the first automation. That automation created the next opportunity.
The doctrine throughout was simple: build a thin layer on top of the harness. Don’t rebuild the harness. We didn’t train a model or write an agent framework. That’s part of why four people could do this at all.
We look at three signals. Tokens per day tell us whether someone is actually using the tool. Pull requests per day tell us whether they’ve figured out agentic work. Business numbers tell us whether any of it is translating into delivery. Each is a gate, not a score. There’s also a fourth signal outside the ladder: when you give someone a skill, do they actually use it?
Are they actually using the tool?
Have they figured out agentic work?
Is it translating into delivery?
When you give someone a skill, do they actually use it?
A hundred PRs a week doesn’t mean a problem needs a hundred PRs.
That person can bring that much force to whichever problem turns out to matter.
We also say capacity, never productivity. A hundred PRs a week doesn’t mean a problem needs a hundred PRs. It means that person can bring that much force to whichever problem turns out to matter.
For comparison, Linear, using the same tools, is at about 13. All three of our engineers sustain 75 or better, and three different people have crossed 100 in three different weeks.
My lead over the team went from 9.5x to about 2x and stayed there for a full quarter while everyone got faster. The gap collapsed because the capability became available to everyone, not because I slowed down.
The tool worked for the person who built it first.
Not because I slowed down.
We’re at v0.8, not v1. “The line builds it and ships it” is the target, and we’re not there yet.
There’s no clean bridge from a single merge to a single customer dollar, and I’m not going to invent one. We also haven’t published a quality series yet. We have the data, and we should publish change-failure rate and review depth.
Every attack that lands on this way of working is a quality attack. The counter isn’t pretending those questions don’t exist. It’s saying what isn’t solved in the same document as what is.
And publish the honest part.
I told a friend no once. He came back a month later with work. His ramp is now the steepest on the team. If I have to convince you, you’re probably not the person. The right people need to be pointed.
We have three seats open: security leader, engineering leader, and VP of Sales. We’re looking for judgment and taste, proven by shipping or by running teams. Have your AI send us what you built in the last 30 days. Competitive pay and equity.
Have your AI send us what you built in the last 30 days.
Chad leads CampusIQ’s move to agentic engineering. Four engineers now merge about 1,150 production PRs a month, roughly the output of a 50-person team, with agents writing most of the code and engineers owning the outcomes, the quality bar, and the calls. He built the machine behind it: a standing merge queue, a fleet of estimation and review bots, executor-validator quality gates, and live telemetry on velocity and AI cost. Last month, more than 20 people across sales, ops, product, and onboarding shipped production code on the same system.