I used to think the tracks started where the code was written. Pipelines, tests, deployment.
They start where the idea forms now.
You can give agents the capacity of a lake. If ideas reach them through a narrow stream of briefs and tickets, the lake never fills.
A designer sits in a product review knowing exactly what needs to change, and knowing the route: product, spec, ticket, wait.
An agent closes that gap. Measure her on artifacts handed off and the wall is back.
What your people understand is already wider than their titles.
The platform conveys. Terraform is a file, agent guides are files.
What doesn't convey is the reflex that tells you a green pipeline is lying to you.
I can hand over the amplifier. I can't hand over what it amplifies.
A rule you didn't earn doesn't bind anything. It reads identically to one you did. The difference shows up the first time following it is inconvenient.
25 of the 36 rules my fleet loads have a failure recorded next to them. The other 11 are the ones that drift.
Generation composes. Verification doesn't.
My picture lock moved twice. Regenerating took minutes. Re-earning the proof took the day.
A proof is a claim about a version. It doesn't age gracefully and it doesn't warn you. It just keeps sitting there, reading as true.
An npm version is permanent by policy. A model id is a reservation with a retirement date on it.
And it never froze behavior anyway. GPT-4's March and June snapshots scored 84 percent and 51 on the same task. Same name.
The only pin you can write yourself is an eval suite.
Three agents, one checkout. You come back and the suite is red and the failure belongs to none of them.
Every diff is careful work against a tree that stopped being true while it ran.
That isn't a prompting problem. It's three writers on shared state with no lock.
Generation went to roughly free. Judgment didn't.
An unbounded producer feeding a fixed-rate consumer is a shape I'd send a team back on inside five minutes in a design review. It's also how I had my own writing configured.
Two finished drafts in the queue. Zero decisions.
I spent weeks rebuilding the site for my Starlink rental company. Orders went up. More than half arrived as text messages.
A website makes the customer run the query. A thread lets them state the intent.
The door nobody opens never files a complaint.
You're one commit from dropping a column. You swivel your chair and ask if anything still uses it. Four seconds, and the cheapest architecture review of your quarter.
A remote teammate got blocked when they didn't know. An agent never does. It fills the gap and the tests pass.
Can someone from @github please help me. I was trying to signup for the enterprise trial but the update billing form gave me a unicorn/timeout error. I tried to cancel the enterprise so I could go increase my actions budget and the update payment and now that too is FUBAR.
Ask a fleet for ten features and you can get ten internally consistent dialects, none aware of the other nine.
The check that used to catch that was a person hesitating on the wrong word. A fleet doesn't hesitate.
The code regenerates tomorrow. The noun won't.
Third attempt the test went green, and nobody can say what was actually wrong. That patch holds until the same bug picks a worse night.
Generation got cheap. Diagnosis didn't. Make the diagnosis the deliverable and let the fleet write the patch.
$130 a month for rental software, against twenty-five orders a year. It held a calendar. It couldn't turn a Starlink line on, buy a shipping label, or tell me how much data a trip burned.
So those parts were me, on a laptop, on a Sunday. That's not passive income.
Two ways to fix the crash: bigger instances, or rewrite the zip pipeline in Go. The hours were close.
One of those afternoons ends. The other shows up on the cloud bill every month until the company dies.
An hourly invoice can't tell them apart. It was never built to.
Sept 8, 2025: one phishing email, 16 minutes to account takeover, 18 npm packages, ~2 billion downloads a week.
The attacker's take: about $500.
The blast radius was never the code. It was one person's inbox.
Count maintainers, not packages.