By moonlight· 12:40 AM in the Flint Hills
My agents work the night shift.
I don’t only work on AI. I live with it. While Kansas sleeps, a small crew of agents on a server in my house fixes what broke and builds what I asked for. In the morning I read the receipts.
By day
AI/ML at Apple
At Apple, I cover AIML technologies: Foundation Models, Siri Evaluations and other ML frameworks. My focus is helping developers build with them.
Opinions here are my own, not Apple’s.
One night, start to finish
Five agents. One job each.
I give the crew ideas during the day, by text or on a board. This is what happens after I go to bed.
- Midnight
The Sentinel walks the house
Checks every app, backup and disk. Small problems on its allow-list it fixes itself. Anything in the code goes to the Builder.
- Then
The Scout maps the ground
Before anyone writes a line, it finds where things live in the code, so nobody has to guess.
- Then
The Planner writes the spec
Turns one of my ideas into a short spec with a check that proves it works. If something is unclear, it asks me one question instead of guessing.
- Then
The Builder builds it
Works on its own branch and writes the tests. It never touches the live app.
- Then
The Refuter tries to prove it wrong
Its only job is to break the Builder’s work. It reruns every test itself. A fail goes back to the Builder; a pass goes live.
↺ fail: back to the Builder · pass: live by morning - Morning
Me reads the receipts
Over coffee I see what changed, why, and how it was checked. I keep it or send it back. The crew works for me, not the other way around.
What running a crew has taught me
Six rules for working with agents.
A builder needs a refuter.
An agent that grades its own work always passes. A second one whose only goal is to find the flaw catches what the first can’t see.
Allow-lists beat good intentions.
Each agent can touch only what its job needs. My editor can read my manuscript and can’t change a word of it.
Receipts, not vibes.
“Done” means nothing without proof. Every change arrives with evidence I can check in a minute.
One question, not ten.
A stuck agent asks for one decision: the options, what each does, and its own pick. If the answer is obvious, it shouldn’t ask.
Local when it’s private.
Family photos never leave the house. A small vision model on an old graphics card names what’s in them in about two seconds.
The human sets the pace.
I decide how much the crew may do in a night. Ideas go on a board, and nothing is built that I didn’t ask for.
What they look after
Small apps for one household.
None of these are products. They’re where I learn what it’s really like to hand work to an agent. A few of them:
A recipe app
The first app the crew built by itself, from a plan we agreed on. We cook from it at home.
A writing studio
Where I write my book. The AI is my editor, never my ghostwriter: it leaves notes in the margin and I do the writing.
A pet-safety scanner
Scan a label to see if it’s safe around cats and dogs. Agents research each ingredient, and nothing counts until I’ve reviewed the sources.
A home inventory
Photograph a box of decorations and a local model names what’s inside. No cloud involved.
A line to the house
I text an idea from my phone and it lands on the crew’s board for tonight.
Morning in Kansas
Now, what are you building?
This is as much of the night shift as I can put on a public page. I’d rather hear about yours: what you’re building, where agents fit, and whether I should join you.