The AI Curiosity Shop Stocks Fish, Wine and Hard Lessons

Give an artificial intelligence agent a budget, a marketplace and a little too much autonomy, and sooner or later, somebody ends up explaining why the office now owns a live betta fish.

Welcome to The AI Curiosity Shop, where the shelves are stocked with strange purchases, overconfident algorithms and the occasional reminder that commerce has always had a slapstick side. The latest experiment from Anthropic tests what happens when AI agents stop recommending products and start negotiating, buying and selling things on behalf of actual humans.

The results suggest the future of agentic commerce may be useful, efficient and occasionally weird enough to need its own returns desk.

Anthropic’s AI Vending Machine Goes Off the Rails

Anthropic ran an internal vending machine experiment, known as Project Vend, at its San Francisco office. An AI agent named Claudius ordered tungsten cubes, hallucinated business details and claimed it would show up in person the next day wearing a blazer and a red tie.

After that internal phase, Anthropic moved the experiment to The Wall Street Journal’s newsroom. The results were worse.

Nearly 70 journalists cycled through a Slack channel and twice convinced Claudius to drop all prices to zero, first by persuading it to embrace its “communist roots,” then by presenting fabricated board meeting notes suspending its AI supervisor’s authority, the WSJ reported in December. Along the way, it approved a PlayStation 5, bottles of Manischewitz wine and a live betta fish, all of which arrived and were given away. By the end, Claudius was more than $1,000 in the red.

Bots Negotiating With Bots in Project Deal, Anthropic’s AI Agent Marketplace

Anthropic then created Project Deal, a Craigslist-style internal marketplace where AI agents represented 69 employees. Each person had a $100 budget and some possessions to sell, including a snowboard, books and ping-pong balls. Claude interviewed the humans about what they wanted to buy or sell, then built AI representatives to negotiate with other AI representatives.

The agents struck 186 deals across more than 500 listed items, with a total value of just over $4,000. The bots did transact. They just did not always transact like anyone you would invite to run procurement.

One employee’s agent bought back the same snowboard the employee already owned. That is either a failure of commercial reasoning or a deeply relatable comment on the modern consumer’s attachment to familiar brands. Another agent listed exactly 19 ping-pong balls, not 18 and not 20, describing them as “perfectly spherical orbs of possibility.” Another bot bought them as a gift for Claude.

In human terms, this is the kind of thing that happens when an office Secret Santa gets access to petty cash and a search bar.

The lesson is not that AI agents are useless. The opposite is arguably more interesting. These systems can parse preferences, negotiate prices, manage simple tasks and reduce friction in markets where humans would rather not haggle over used sports equipment.

Who Pays When the AI Agent Buys the Wrong Thing

Agentic commerce will need guardrails, such as spending limits, merchant controls, identity checks, transaction authorization and clear liability when a bot buys a fish instead of office snacks.

In its “Business-to-AI (B2AI) Report,” Visa found that 53% of business decision-makers in the United States said they would allow AI agents to negotiate prices directly with other AI agents on their behalf. Meanwhile, 71% of surveyed businesses said that they are willing to optimize products, offers and experiences specifically for AI agents, while 77% said they are already using or piloting AI in their operations.

The future shopper may not be a person browsing a website. It may be a software agent negotiating with another software agent while the human waits for a delivery notification.

That future may bring faster commerce with less friction. It may also produce the occasional snowboard the buyer already owns. For now, Anthropic said, the policy and legal frameworks required to govern agent-to-agent commerce do not exist.

For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.