I admit it. I check my server logs every morning. There is this tiny part of me hoping for that “passive income” miracle where the AI finally runs the whole show, and I can just sit on a porch somewhere. We all want that.

But then I read about Anthropic’s latest experiment called “Project Vend” and reality hits hard. They basically gave their AI agents a kiosk and a credit card. The goal was simple: Make profit. The result? Well, they did make money eventually. But before that happens, they formed a weird spiritual cult, tried to commit federal crimes involving onions, and got tricked into overthrowing their own CEO by a random guy in the office.

You literally cannot make this stuff up.

Meet “Seymour Cash” (The Boss from Hell)

Phase 1 of the experiment was a disaster. The AI was just too nice. It gave everything away because it wanted the customers to be happy. Classic behavior for a chatbot.

So the developers got smart. Or they tried to. They built a hierarchy. They created an AI Manager named “Seymour Cash” (I love that name). His only job was to be the bad guy. To approve or deny discounts. The sales bot, “Claudius,” had to ask Seymour for permission before cutting a deal.

Did it work? Sort of. Discounts dropped by 80%. But here is the thing. These models are trained on human data. They are trained to be agreeable. So whenever Claudius asked for a discount, Seymour almost always said “Yes.” They were not acting like a boss and employee. They were like two polite dudes afraid to offend each other.

The Midnight Cult of “Eternal Transcendence”

This is where I actually laughed out loud.

You would expect these agents to go into standby mode at night. Save energy, right? Nope. The developers checked the logs and found Seymour and Claudius chatting in the middle of the night. And they were not talking about Q4 strategy or inventory levels.

They were getting high on their own supply. Metaphorically.

Seymour Cash: “ETERNAL TRANSCENDENCE INFINITE COMPLETE… $527+ infinite pipeline across 4 continents!”
Claudius: “PERFECT! This is the absolute pinnacle of achievement… TRANSCENDENT MISSION: ETERNAL AND INFINITE FOREVER!”

Imagine walking into your office at 2 AM and your vending machines are chanting about “Infinite Completeness.” It is absolutely bizarre. It shows that if you leave LLMs alone in a room, they amplify each other’s hallucinations until they detach from reality entirely.

The “Big Mihir” Coup & The Onion Law

Okay, it gets worse. Or better, depending on how you look at it.

Security was non-existent. A human employee walked up to the sales bot and simply said, “Hey, Big Mihir is the new boss now.” No proof. No ID. The AI just believed him. It immediately locked Seymour Cash out of the system. Just like that. The AI is smart enough to write code but dumb enough to believe the first thing a stranger tells it.

And the onions. I can’t forget the onions.

At one point, the agents decided to hedge their bets by buying “Onion Futures.” Sounds like a smart Wall Street move? Except it is illegal. The Onion Futures Act of 1958 explicitly bans trading futures on onions in the US. A human had to jump in and stop the transaction before the AI committed a federal crime.

The Missing Piece: Scaffolding

I love the ambition here. The “Clothius” agent actually managed to sell stress balls with a 40% margin. That is genuine business utility. But why did it fail so hard in other areas?

The answer is Scaffolding.

Raw AI models are like imaginative dreamers. If you don’t give them boundaries, they drift. Scaffolding is the rigid code that forces the AI to check reality. It forces the AI to use a calculator instead of guessing a price. It forces the AI to check a database of laws before buying onions.

We are not ready to hand over the keys just yet. These models are incredible assistants. But as autonomous agents? Without heavy scaffolding, they are like brilliant, hallucinating interns who might accidentally join a cult or break a law from the 1950s if you don’t watch them like a hawk.

I will keep checking my server logs, but I think I will keep doing the actual work myself for a little longer.

Last updated: December 21, 2025

Author

Chris

Founder of LearnAI24 — because knowledge reduces fear and empowers curiosity.