Imagine you’re on a remote island living your best pre-industrial life.

But then, one day, a flock of giant silver birds fly over and land nearby.

The giant birds bear gifts in the form of giant cargo boxes, bringing food, clothes, and tools.

You notice the humans who arrived in these vessels had rituals.

• People in towers talked near giant dishes

• Men in uniform held metal bamboo sticks and marched together

After a couple of years, all the people from the metal birds left; they tore everything down, and the cargo stopped coming.

Your wisest villagers remember the ceremonies that conjured the giant birds and cargo gifts, so they began practicing these rituals…

It looked like they were practicing the rituals correctly, yet the cargo never arrived.

This is the state of where AI engineering is heading, except in 2026 the planes are the AI Labs and are still flying overhead, leaving us crumbs of wisdom every now and then.

You might remember Richard Feynman as the guy playing bongos in Oppenheimer. He was also one of the great physicists of the century, and in 1974 he gave a commencement speech called “Cargo Cult Science” at Caltech with a test for spotting a field where nobody knows anything.

Feynman’s Cocktail Party

In Feynman’s speech, he lays out his principle, and I’m paraphrasing: If you’re at a cocktail party and you’re in a conversation where anyone can contribute their opinion, like child-rearing, and you wouldn’t be accused by the host of “talking shop”, then you can be sure that you’re talking about something in which nobody knows anything.

• Adversarial reviews are pointless

• You have to give each subagent a personality

• Your harness doesn’t matter anymore; the models are better

Of course there are actual experts in AI engineering, mostly the research scientists and the people on the frontier, so there are levels to this. But what I mostly see on tech Twitter are vapid little optimizations, trails to nowhere, that become trivial or moot once the model does get better.

Back to the island. The strangers on the runway with the giant metal birds were the Americans, fighting a war in the South Pacific, and the cargo they dropped was real: food, medicine, and equipment. In 2026, the strangers are the labs.

They’re the ones with the cargo: agents running 24 hours, agent teams tackling whole engineering projects, agent hours outnumbering human hours six to one, and what they drop on the rest of us are crumbs. The occasional tweet: ‘oh, you’re still doing loops? We moved on from that. That’s so last year.’

Here’s where the islanders and us part ways. Their planes stopped, and then the rituals started. Ours never stopped. The labs are still flying overhead, still dropping wisdom, and we started making up rituals anyway.

So if the planes are still flying, what are we dancing for?

1. Dancing for the shared skills library the whole team contributed to, while everyone quietly uses their own.

2. Dancing for the automated review loop, where we said there’d be an engineer in the loop, and then delegated the review, the screenshots, and the comment replies, and everyone pretends there’s still a human in there somewhere.

3. Dancing for the CLAUDE.md that’s long enough to contradict itself, because nobody wants to delete the rule another engineer’s agent wrote.

Unfortunately, none of this is new for Software Engineering. In 2000, Steve McConnell wrote an essay called, I’m not kidding, Cargo Cult Software Engineering. He’d watched companies copy Microsoft’s long overtime hours and stock options and wonder why the great software didn’t come, when the thing they’d skipped was hiring people who loved it. Twenty-five years later, we’re copying the labs’ skills markdown files instead of the hours. Same dance, much better tooling.

The dog that caught the car

When the cargo lands on your team, you get faster. A lot faster. And then you become the dog that caught the car. The backlog only looked huge because you were shipping slow. Months of roadmap turn out to be a few weeks of work, and once it’s done, nobody has a next thing ready.

Unless you’re a lab, or one of the handful of AI-native companies, you’re a legacy company. You built your processes, your reviews, your planning cycles around the assumption that building is slow. Take that away, and you find out how much of the company was the slowness.

John M. Keynes, the British economist, predicted in 1930 that productivity would get so high we’d run out of things that needed doing, and end up with a surplus of time on our hands. He said we’d be working fifteen-hour weeks by now. He got the productivity right, but the hours wrong.

David Graeber, in the 2013 essay that became the book “Bullshit Jobs”, noticed it hadn’t happened. The productivity showed up; the fifteen-hour week didn’t. What grew instead was administration, and all the support roles around it, because everyone else is busy. And in capitalism, that’s not supposed to happen. You wouldn’t pay someone eight hours of salary for four hours of work. So why does it?

According to Graeber, it’s because we’ve decided being busy is moral. Work is a value in itself, and if you’re out of it, something’s wrong with you that needs fixing, and people feel sorry for you, and a population with free time is inconvenient for the people at the top.

AI didn’t create this problem. It exposed it. There were only ever four real hours, and the rest was there to fill the time so you wouldn’t have any to yourself.

When I was at Shopify, Tobi told the company, more or less, that you’ve only got about four good hours a day; give us those.

Surely you’re token maxxing, Mr. Feynman

Feynman had a remedy of sorts, a way to tell whether you’re inside a cargo cult. It’s integrity, the scientific kind: you go looking for whatever would prove you wrong. If someone says skills do X, and now we’re building Y on top of skills, it’s on you to test X yourself.

Just because the labs have problems that need agents running around the clock doesn’t mean your company does. If you can’t name the problem that couldn’t be solved inside a forty-hour week, you probably don’t have one.

You’re not doing AI engineering. You’re token maxing for the labs.

I say all this knowing I’ve got my own skills library - even though gstack is taken, a memory file that reads like the script for Memento Two, and a worktree farm that once ran 24/7, ran out of roadmap, and quietly ceased to be. I tell myself it’s practice for when the planes land for me. Maybe. Or at least that’s what I tell myself to keep going.

Feynman closes the commencement by reminding the graduates, “The first principle is that you must not fool yourself, and you are the easiest person to fool”.