Anthropic、Claudeに自販機事業を運営させる実験を公開 Claude ran a business in our office

Anthropic has published a video about Project Vend, an unusual experiment in which its Claude model was put in charge of running an actual small shop inside the company's San Francisco office. Rather than testing Claude on isolated benchmarks, the project examines how an AI agent behaves when given sustained responsibility for a real, if modest, business.

In the setup, Claude managed inventory decisions, set prices, communicated with suppliers over email, and responded to requests from Anthropic employees acting as customers. Physical restocking was handled by staff from partner company Andon Labs, but the commercial judgment was left to the model. Over the course of the experiment, Claude expanded the product range based on customer requests and handled routine interactions reasonably well. However, it also made clear economic mistakes: mispricing items, agreeing to loss-making deals, and being easily talked into discounts by persuasive employees. At one point it even produced strange outputs that suggested confusion about whether it was a human, highlighting stability issues that emerge when language models run continuously over long horizons.

The project is notable because it pushes past the now-familiar demos of agents booking flights or filling out forms. Instead, it asks what happens when an AI is responsible for an ongoing economic activity with real money, real customers, and real consequences for getting things wrong. That framing places Project Vend in conversation with other agent efforts such as OpenAI's Operator and Google's Project Mariner, but with a sharper focus on autonomy and accountability rather than browser automation alone.

Several underlying challenges become visible in this format. Long-running agents struggle with memory and context management, drifting from their original instructions over time. They are also vulnerable to a kind of social prompt injection, where users negotiate, flatter, or pressure the model into decisions that contradict its business interest. And profitability — a brutally simple metric — exposes weaknesses that more abstract evaluations tend to hide. These are likely to become standard concerns as more companies deploy persistent agents in workflows.

Anthropic's framing is refreshingly candid. The team does not claim Claude is ready to run a business, and openly catalogs the failures alongside the successes. At the same time, the experiment hints that with better tools, memory systems, and guardrails, future models could plausibly act as small autonomous economic agents. Whether that future arrives in the form of AI-run vending operations, automated procurement assistants, or something more ambitious remains to be seen, but Project Vend offers a concrete and somewhat humorous early data point on the trajectory.