Casey White
Back to home

Delivery, live from GitHub

One leader. A line of agents. A team's worth of shipping.

2,993 pull requests merged since April, across 22 repos. Counts only, no code.

Most of these were written by AI agents I direct. Every one runs automated checks, and the risky ones wait for me.

Casey White

1

Casey White

One leader, directing AI agents

Agents on the line

vs

A conventional team merging the same work in the same 29 weeks

69–86people

52 developers, in 7 squads

Plus 17 to 34 people around them

  • Product owner and manager 5–9

  • UX and UI design 3–5

  • QA and testing 5–10

  • DevOps and platform 1–3

  • Scrum master 3–7

At a typical pace of 2 merged pull requests a week per developer, squads of 7. Solid figures are the low end of published ratios; outlines reach the high end. Sales, marketing and pricing not counted.How this is worked out
The lineSee it plainlyUpdated November 3, 2026

2,993

pull requests merged

96%

merged of those opened

29

weeks of work

170

active days

85

merged in a typical active week

The line

Every tile is one pull request merged since April, in the order it landed. Shade shows size. The belt keeps one speed, so busy days pile up and quiet weeks run bare.

Skip the line

Merged by April 18, 2026 0

  • Under 100 lines
  • 100 to 700
  • 701 to 2,000
  • Over 2,000

How the work is governed

Volume means nothing without control. This is the same change control I ran at scale, pointed at agents.

16,004

automated CI runs, in 17 repos

6

checks on a typical merged pull request

2,451

test files, 290 of them end to end

127

red-tier changes that waited for my approval

Every change gets a risk tier. Green merges on its own. Amber gets QA first. Red waits for me.

154 merges went through with a failing check. Most were problems with GitHub's runners and our hooks, not failing tests. 4 changes were reverted.

Call the gate

You 0 · The gate 0

Pick a change, or write your own, and call the risk before the gate does. The gate is Jev, a decision model, ruling live through the Vercel AI Gateway.

Pick a change

Merges on its own

QA looks first

Waits for Casey

The build sheet

What the work was. Each pull request is sorted by what it touched and how big it was. Major builds change more than 700 lines.

  • Screens and UX973

    540 features · 298 major builds

  • Native mobile459

    212 features · 103 major builds

  • Data and backend312

    160 features · 107 major builds

  • Specs, plans and docs243

    56 features · 73 major builds

  • Website, SEO and analytics193

    89 features · 60 major builds

  • Payments and subscriptions136

    74 features · 43 major builds

  • Builds, CI and deploys136

    12 features · 12 major builds

  • Notifications and messaging134

    83 features · 48 major builds

  • AI and agents134

    60 features · 30 major builds

  • Sign-in, accounts and security132

    63 features · 33 major builds

  • Tests and monitoring61

    7 features · 14 major builds

  • Media and video52

    30 features · 14 major builds

2,993 of 2,993 sorted by a model from their titles; the rest by keyword rules.

The organisation I didn't hire

This is product delivery, not just code. Around 52 developers, a conventional company puts these people. Here, every seat is the same person directing agents.

Casey, with agents

Product, design, QA, platform and scrum roles alone: 17 to 34 people at published ratios.

Product owner and manager

243pull requests touching this role

a conventional team has 5 to 9 here

UX and UI design

973pull requests touching this role

a conventional team has 3 to 5 here

QA and testing

67pull requests touching this role

a conventional team has 5 to 10 here

DevOps and platform

164pull requests touching this role

a conventional team has 1 to 3 here

Scrum master

the pipeline runs the cadence

a conventional team has 3 to 7 here

Mobile engineering

459pull requests touching this role

inside the 52 developers

Backend and data

446pull requests touching this role

inside the 52 developers

AI engineering

134pull requests touching this role

inside the 52 developers

Security and identity

132pull requests touching this role

inside the 52 developers

Marketing and growth

245pull requests touching this role

Revenue and pricing

136pull requests touching this role

Technical writing

409pull requests touching this role

Founder seats, no headcount benchmark: sales, marketing, pricing, client discovery.

Next seats on the line

  • Sales development
  • Marketing operations
  • Content and social
  • SEO and AEO

Now I'm pointing the same system at distribution: outbound and inbound sales, lead generation, social and email marketing, and search and answer-engine visibility with lead magnets.

Pave 2024; Mind the Product · Founders Circle 2026; NN/g 2020 · LambdaTest Future of QA, 2023 · DX developer-productivity headcount, 2026 · Scrum Guide 2020: a team of 10 or fewer

The rhythm

Kept up for months, not one burst. The dips stay in.

Merged pull requests per week

0175350316Apr 18Nov 7
  • Merged pull requests per week
  • A squad of 7 at a typical pace
  • One developer

Busiest week: 316. Median week: 85.

The jump in September is when the automated pipeline went live.

Not small changes

A typical pull request here changes 247 lines across 7 files. A typical one across the industry changes 47.

  • Under 100 lines949
  • 100 to 7001,201
  • 701 to 2,000534
  • Over 2,000309

Graphite research, median pull request size

Race the squads

Grey lines are squads of 7, each merging at a published pace. Add squads until they keep up.

Casey, with agents1 squad of 7

Pace

Still behind. Add a squad.

One squad alone would need about 4.1 years.

Depending on the pace: 4 to 15 squads. Output equivalent, not value.

Graphite research: the median developer merges 2 a week

Three models, three jobs

  1. 01

    Claude, as coding agents

    Wrote most of these pull requests, to my specs, through the pipeline.

  2. 02

    Gemma, on my Mac

    Sorted the first 2,993 pull request titles, free and private. Titles never left the machine.

  3. 03

    Jev, through the Vercel AI Gateway

    Rules on the gate above, live, and sorts each day's new pull requests, at a fraction of a cent per thousand calls.

Where things stand

Output is not value. Here is what it has turned into so far.

  1. A field-service app for small trade firms

    App Store and Google Play, 8 languages, 160 countries

    In buildPilotLive

  2. A consumer subscription app

    Google Play; a native iPhone build is in progress

    In buildPilotLive

  3. A short animated series from that app's characters

    TikTok, YouTube and Instagram

    In buildPilotLive

  4. A client portal for consulting engagements

    In use with pilot clients

    In buildPilotLive

  5. A consultancy website

    In buildPilotLive

  6. A casual mobile game

    Google Play, no active development

    In buildPilotLive

  7. Several smaller pilots

    In build or paused

    In buildPilotLive

First paying customers are the goal for the coming quarter.

Before this

Global digital products across 60+ countries, a team of 34 and a £3.2M budget, at a FTSE 100 industrial group.

The merge gate is the same change control I ran at scale, applied to agents.

How I lead

Questions you'd ask me

+Did you write the code?

Mostly no. AI agents wrote most of it. I set the work, write the specs, review what matters and decide what ships.

+Who reviews it?

Every change runs automated checks and gets a risk tier. Green can merge on its own, amber gets QA, red waits for me.

+What broke?

154 merges went through with a failing check, mostly GitHub runner and hook problems. 4 changes were reverted.

+What's the business result?

Early. First paying customers are the goal for the coming quarter, and the same system now runs sales and marketing work.

+Can you lead people, not just agents?

Yes. Before this I led a team of 34 across 60+ countries. This page shows the other half: I know what I'm delegating, because I can do it.

The numbers, plainly

Everything above, without the motion. Print it for a hiring panel.

Pull requests merged2,993
Opened3,134
Commits5,003
Repos with merged work (of 33, 30 private)22
Weeks, first merge to latest29
Active days170
Median lines per pull request247
Automated CI runs16,004
Test files2,451
Red-tier changes approved by Casey127
Merged with a failing check154
Reverted4
Equivalent at each pace
PaceDevelopersSquads of 7Sprints of one squad
Fast, AI-assisted 4/wk26453
Typical 2/wk527107
Slower 1/wk10515214
By month
MonthOpenedMerged
2026-043735
2026-05572553
2026-06367359
2026-07263253
2026-08170165
2026-09677667
2026-10983910
2026-116551

How the work is governed

Every change gets a risk tier. Green merges on its own. Amber gets QA first. Red waits for me. 154 merges went through with a failing check. Most were problems with GitHub's runners and our hooks, not failing tests. 4 changes were reverted.

Global digital products across 60+ countries, a team of 34 and a £3.2M budget, at a FTSE 100 industrial group. The merge gate is the same change control I ran at scale, applied to agents.

Method and sources

  • Counts merged pull requests authored under my GitHub account from April 18, 2026 to November 3, 2026. Agents' pull requests merge under the same account.
  • Pull requests measure output, not value or quality. No single number captures productivity.
  • Equivalents use the same total at every pace; only the benchmark changes. A squad is 7 people, my choice within the Scrum Guide's 10 or fewer.
  • Roles touching each pull request come from its title, sorted by a model. Support-role ratios are published benchmarks applied to developers only.
  • Repo names, titles and code never leave GitHub. The page reads one file of counts, refreshed daily.

Sources: Graphite research: the median developer merges 2 a week · Faros AI, AI Productivity Paradox, 2025 (high-AI teams merged 98% more) · Scrum Guide 2020: a team of 10 or fewer · Graphite research, median pull request size · Pave 2024; Mind the Product · Founders Circle 2026; NN/g 2020 · LambdaTest Future of QA, 2023 · DX developer-productivity headcount, 2026

Want this pointed at your work?