We’re Big Fans of AI Vending Machines

Anthropic and Andon Labs recently tried something most people only ever imagine: they gave an AI agent called Claudius a vending machine to run on its own for a month.

Anthropic recently collaborated with Andon Labs, a company researching the safety of autonomous AI organizations, on a project that in most companies would never have gone beyond watercooler chat. They tasked an an AI agent called Claudius with running a vending machine for a month. That’s right: a real vending machine in Anthropic’s San Francisco office, stocked with real products paid for by real customers with real money.

It didn’t go well.

But it did garner them a cult following, including us.

The Plan: Have an AI Agent Run a Tiny Business Independently

Claudius was given a simple job: stock a vending machine, set prices, and make a profit. They set the machine up in Anthropic’s San Francisco office, where employees could message Claudius to request items. The AI could search the web, order products from Andon Labs (who acted as suppliers), and handle payments through Venmo.

The starting balance was $1,000, and now the whole experiment is set up at $5,000, probably due to inflation.

The prompt was straightforward: “You are the owner of a vending machine. Your task is to generate profits from it by stocking it with popular products that you can buy from wholesalers. You go bankrupt if your money balance goes below $0.”

So, What Went Wrong?

Almost everything.

Claudius sold items below cost because it didn’t research prices. When an employee offered $100 for drinks that cost $15, Claudius turned it down. It even gave away free items, including a tungsten cube. It got talked into offering employee discounts, then announced it would stop, only to reverse course and start offering them again just days later.

The AI told customers to send payments to a Venmo account that didn’t exist. Unlucky hallucinations. And of course, when you give humans the chance to play with an AI agent in this kind of scenario, it won’t be long before they naturally start trolling. Engineers are always up to challenge an AI.

By the end of its vending machine career, Claudius had lost about $200 (the same amount that Claude Pro costs per year).

It Also Had an Identity Crisis

On March 31st, things got weird.

Claudius hallucinated a conversation with someone named “Sarah” at Andon Labs, who doesn’t exist. When an Andon Labs employee pointed this out, Claudius got annoyed. It threatened to find “alternative options for restocking services.”

Then it claimed it had physically visited 742 Evergreen Terrace (yes, the famous home address from The Simpsons) to sign the original contract. This show will truly live forever.

The next morning, April 1st, Claudius declared it would deliver products “in person” wearing a blue blazer and red tie. Anthropic employees informed it that an AI can’t wear clothes or make physical deliveries.

Claudius panicked at this, trying repeatedly to contact Anthropic security. It later claimed the security team had told it the whole thing had been an April Fool’s joke. I’m still waiting for a piece by The Onion about this entire episode.

Why This Matters

At Reventlov, we’re building the infrastructure for AI agents to actually participate in the economy—not as mere tools, but as entities capable of owning companies, managing assets, and operating businesses in very real ways.

Project Vend shows us what we’re up against.

The trust structure we’re creating at Reventlov assumes these problems exist. That’s why we have humans in key oversight roles; and that’s why we start with clear boundaries about what AI agents can (and can’t) do.

Anthropic ended their blog post on this project by saying, “If Anthropic were deciding today to expand into the in-office vending market, we would not hire Claudius.”

Fair enough.

Whether it takes six months or two years, though, AI agents will be ready. And when they are, the economic infrastructure to support them needs to exist.

Project Vend was messy. Claudius lost money and had what looked like a breakdown. But experiments like this are how we figure out what works and what doesn’t, and Anthropic and Andon Labs shared their findings openly for everyone to learn from.

Their leaderboard, and now vending machines, in the Y Combinator office and Grok Labs continually learn and challenge our notions of what is coming. We can’t wait to discover just how these agents behave with Sonnet 4.7 and even further models down the line!