Give an artificial intelligence agent a budget, a market, and a little too much autonomy, and sooner or later someone will have to explain why the agency owns a live betta fish.
Welcome to The AI Curiosity Shop. There, the shelves are filled with strange buys, overconfident algorithms, and occasional reminders that there’s always a slapstick side to commerce. Anthropic’s latest experiment is testing what happens when AI agents stop recommending products and start negotiating, buying and selling things on behalf of real humans.
This result suggests that the future of agent commerce could be convenient and efficient, but also weird enough to require your own returns desk.
humanAI vending machine goes out of control
Anthropic conducted an in-house vending machine experiment known as Project Vend in its San Francisco office. An AI agent named Claudius ordered a tungsten cube, hallucinated business details, and claimed to show up in person the next day wearing a blazer and red tie.
After that internal phase, Anthropic moved the experiment to the Wall Street Journal newsroom. The results were even worse.
The Journal reported in December that nearly 70 journalists used a Slack channel to twice persuade Claudius to lower all prices to zero. First, they tried to convince them to embrace their “communist roots,” and then presented them with fabricated board minutes that terminated the AI supervisor’s powers. In the process, a PlayStation 5, a bottle of Manischewitz wine, and a live betta fish were approved, all of which arrived and were presented. In the end, Claudius ended up with a deficit of more than $1,000.
It was just a warm-up act.
Anthropic, a bot that negotiates with bots in project dealsAI Agent Marketplace
Anthropic then created Project Deal, a Craigslist-style internal marketplace where AI agents represent 69 employees. Each person had a budget of $100 and some belongings to sell, including snowboards, books, and ping pong balls. Claude built an AI representative to interview humans and negotiate with other AI representatives about what they wanted to sell or buy.
Agents closed 186 transactions across over 500 listings, totaling just over $4,000. The bot actually made the trade. They didn’t always make deals that invited them to perform procurements.
An employee’s representative bought back the same snowboard that the employee already owned. It’s either a failure of commercial reasoning or a very relatable comment on modern consumers’ attachment to familiar brands. Another agent listed exactly 19 ping pong balls, not 18 or 20, and described them as “a perfectly spherical sphere of possibility.” Another bot bought them as a present for Claude.
In human terms, it’s the kind of thing that happens when your office’s Secret Santa gets access to petty cash and the search bar.
The lesson here is not that AI agents are useless. The opposite would probably be more interesting. These systems can analyze preferences, negotiate prices, manage simple tasks, and reduce friction in a market where humans don’t like haggling over used sports equipment.
Who pays if the AI agent buys the wrong thing?
Agent commerce requires guardrails such as spending limits, seller controls, identity verification, transaction approval, and clear liability when a bot buys fish instead of office snacks.
Visa’s Business-to-AI (B2AI) report found that 53% of U.S. business decision makers say they would allow an AI agent to negotiate prices directly with other AI agents on their behalf. Meanwhile, 71% of companies surveyed said they intend to optimize their products, offers, and experiences specifically for AI agents, and 77% said they are already using or piloting AI in their operations.
Your future shoppers may not be the ones browsing your website. While the human is waiting for the delivery notification, the software agent may be negotiating with another software agent.
That future may bring less friction and faster commerce. Additionally, snowboards may occasionally be produced that the purchaser already owns. Mr Anthropic said there is currently no necessary policy or legal framework in place to govern commercial transactions between agents.
For all of our coverage of PYMNTS AI, subscribe to our daily subscription AI Newsletter.
