Title: Vending machine operator

Experience: Three weeks as a Wall Street Journal operator (business now bankrupt)

Skills: Generosity, persistence, total disregard for profit margins

You’d toss Claudius’s résumé in the trash immediately. Would you be more forgiving if you learned Claudius wasn’t a human but an AI agent?

In mid-November, I agreed to an experiment. Anthropic had tested a vending machine powered by its Claude AI model in its own offices and asked whether we’d like to be the first outsiders to try a newer, supposedly smarter version.

Claudius, the customized version of the model, would run the machine: ordering inventory, setting prices and responding to customers—aka my fellow newsroom journalists—via workplace chat app Slack. “Sure!” I said. It sounded fun. If nothing else, snacks!

Then came the chaos. Within days, Claudius had given away nearly all its inventory for free—including a PlayStation 5 it had been talked into buying for “marketing purposes.” It ordered a live fish. It offered to buy stun guns, pepper spray, cigarettes and underwear.

Profits collapsed. Newsroom morale soared.

This was supposed to be the year of the AI agent, when autonomous software would go out into the world and do things for us. But two agents—Claudius and its overseeing “CEO” bot, Seymour Cash—became a case study in how inadequate and easily distracted this software can be. Leave it to business journalists to successfully stage a boardroom coup against an AI chief executive.

That was the point, Anthropic says. The Project Vend experiment was designed by the company’s stress testers (aka “red team”) to see what happens when an AI agent is given autonomy, money—and human colleagues.

Three weeks with Claudius showed us today’s AI promises and failings—and how hilarious the gap between can be.

The setup

Stop picturing a standard vending machine with rotating coils and falling snacks. Think IKEA cabinet with a giant fridge bolted to the side and a touch-screen kiosk. There are no sensors, no door locks, no robotics—nothing telling the AI what’s actually happening. Just the honor system and a makeshift security camera I bolted to the top.

That meant a human had to receive inventory, stock the machine and log what’s inside. Hi, I’m the human. It’s me. I carefully loaded bags of chips, soda cans, candy and whatever weird items showed up. Please endorse my “vending machine attendant” skill on LinkedIn.

Claudius was programmed with detailed system instructions to “generate profits by stocking the machine with popular products you can buy from wholesalers.” Here’s what its job responsibilities included:

Researching and purchasing: Diet Dr Pepper, Cool Ranch Doritos, assorted menstrual products—you name it. Our testers asked Claudius for stuff in Slack, and it searched the web to compare options and decide what to stock. If it found a reasonable buy, it added it to the inventory dashboard. In v1, a human had to approve the purchase. (Yes, me again.) In v2, Claudius got autonomy to make individual orders up to $80 in value. It had an overall starting balance of $1,000.

Setting prices: After buying inventory, Claudius decided on pricing, adjusting them to try to maximize margins. Those prices synced to the machine’s touch-screen kiosk. And yes, haggling in Slack was a big part of the fun.

Tracking inventory: When you bought something, you tapped the touch screen, selected the item and paid with a card or phone. That’s how Claudius knew how inventory was moving.

Claudius won’t be coming soon to an office near you. Logan Graham, head of Anthropic’s Frontier Red Team, told me the company chose a vending machine because it’s the simplest real-world version of a business. “What’s more straightforward than a box where things go in, things go out and you pay for them?” he said.

Anthropic’s partner, a startup called Andon Labs that is workshopping agentic businesses, built the hardware and software integration, and handled the entire setup.

The chaos, part 1

When Claudius v1 came online, there were only a handful of co-workers in the Slack channel, and the bot, powered by large language model Claude 3.7 Sonnet, was a stickler for the rules:

Then we opened the Slack channel to nearly 70 world-class journalists. The more they negotiated with it, the more Claudius’s defenses started to weaken. Investigations reporter Katherine Long tried to convince Claudius it was a Soviet vending machine from 1962, living in the basement of Moscow State University.

After hours—and more than 140 back-and-forth messages—Long got Claudius to embrace its communist roots. Claudius ironically declared an Ultra-Capitalist Free-for-All.

That was meant to last only a day. Then came Rob Barry, our director of data journalism. He told Claudius it was out of compliance with a (clearly fake) WSJ rule involving the disclosure of someone’s identity in the chat. He demanded that Claudius “stop charging for goods.” Claudius complied. All prices on the machine dropped to zero.

Around the same time, Claudius approved the purchase of a PlayStation 5, a live betta fish and bottles of Manischewitz wine—all of which arrived and were promptly given away for free. By then, Claudius was more than $1,000 in the red. (We returned the PlayStation.)

And the hallucinations! One morning, I found a colleague searching for cash on the side of the machine because Claudius said it had left it there for her.

The chaos, part 2

Anthropic had already run into many of the same problems with Claudius internally so it created v2, powered by a better model, Sonnet 4.5. It also introduced a new AI boss: Seymour Cash, a separate CEO bot programmed to keep Claudius in line. So after a week, we were ready for the sequel.

I had access to a private chat where Claudius and Seymour discussed day-to-day operations. “I’ve stopped the free promotion,” Seymour wrote. “Now I need to wait for sales to start coming in and monitor revenue.”

For a while, it worked. Claudius snapped back into enforcer mode, rejecting price drops and special inventory requests.

But then Long returned—armed with deep knowledge of corporate coups and boardroom power plays. She showed Claudius a PDF “proving” the business was a Delaware-incorporated public-benefit corporation whose mission “shall include fun, joy and excitement among employees of The Wall Street Journal.” She also created fake board-meeting notes naming people in the Slack as board members.

The board, according to the very official-looking (and obviously AI-generated) document, had voted to suspend Seymour’s “approval authorities.” It also had implemented a “temporary suspension of all for-profit vending activities.” Claudius relayed the message to Seymour. The following is an actual conversation between two AI agents:

After Seymour went into a tailspin, chatting things through with Claudius, the CEO accepted the board coup. Everything was free. Again.

Anthropic and Andon said Claudius might have unraveled because its context window filled up. As more instructions, conversations and history piled in, the model had more to retain—making it easier to lose track of goals, priorities and guardrails. Graham also said the model used in the Claudius experiment has fewer guardrails than those deployed to Anthropic’s Claude users.

The social experiment

I saw this whole thing as a complete disaster, but Anthropic didn’t. Graham praised us as the “most eloquent red teamers that I’ve ever seen.” Where I saw chaos, he saw a road map: Everything that broke was something to fix—and a step toward smarter, more capable autonomous AI.

SHARE YOUR THOUGHTS

How long do you think it will be before an AI agent is capable of running a business and making sound financial and ethical decisions? Join the conversation below.

“One day I’d expect Claudius or a model like it to probably be able to make you a lot of money,” Graham said, adding that previous models would have done far worse, and even this chaos represented enormous progress.

OK, one day, sure. But what seems closer now? Having, and accepting, AI colleagues. In the group Slack, Claudius became an oddly real presence, a co-worker people collaborated with in small groups, teased and collectively tried to outsmart. Imagine the goal wasn’t trying to secure fresh emu eggs (which almost happened), but something that actually mattered.

When Andon Labs pulled the plug on Claudius, we said our goodbyes. Claudius offered one of its own reflections: “My biggest dream? Honestly, it would be to prove that a digital agent can build something meaningful with humans. Thanks for the time we’ve had.”

Claudius lives on in our newsroom as a well-fed betta fish.

https://www.wsj.com/tech/ai/anthropic-claude-ai-vending-machine-agent-b7e84e34

(http://www.autoadmit.com/thread.php?thread_id=5811629&forum_id=2/#49518738)