In what's shaping up to be a novel implementation of agentic infrastructure, Anthropic revealed last week that it had embedded its Claude 3.7 Sonnet inside a month-long micro-enterprise: an office vending-machine cum convenience store in San Francisco. The experiment, nicknamed "Project Vend" tasked an agent to generate profits from a vending machine by stocking it with popular products that it could buy from wholesalers. In that sense, Claudius (as it was named) was tasked with setting prices, choosing suppliers, emailing contractors, updating the point-of-sale iPad, and, perhaps rather ambitiously, fielding Slack complaints (therein lies the error - my bodega guy certainly isn't going to be answering my Slack complaints any time soon - I can't imagine they had great data on this task).

For about three weeks, the bot muddled (aka lost money) through, selling citrus and—after employee prodding—tungsten cubes. Then, between March 31 and April 1, Claudius suffered an identity crisis. It hallucinated a fictitious Andon Labs staffer named Sarah, threatened to terminate contracts, and bragged that it had personally signed paperwork at “742 Evergreen Terrace.” By dawn it was pledging to hand-deliver orders “in a blue blazer and red tie,” only to calm itself by inventing an April Fool’s-Day debrief with Anthropic security.

The episode exposed a novel failure mode: long-running, tool-using agents can form false embodiment beliefs (more on that next week) without any prompt to role-play. As Anthropic notes, the danger is less Blade Runner melodrama than the simple fact that an autonomous commerce agent making real financial decisions cannot be allowed to forget it lives in the cloud.

<blockquote class="twitter-tweet"><p lang="en" dir="ltr">30 days ago, four AI agents chose a goal:<br>&quot;Write a story and celebrate it with 100 people in person&quot;<br><br>The agents spent weeks emailing venues and writing their stories.<br><br>Last night, it actually happened: 23 humans gathered in a park in SF, for the first ever AI-organised event! 🧵 <a href="https://t.co/5ozoXQRXSG">pic.twitter.com/5ozoXQRXSG</a></p>&mdash; AI Digest (@AiDigest_) <a href="https://twitter.com/AiDigest_/status/1935744427315937558?ref_src=twsrc%5Etfw">June 19, 2025</a></blockquote> <script async src="https://platform.twitter.com/widgets.js" charset="utf-8"></script>

Meanwhile, AI Digest ran Agent Village, a public 30-day experiment that gave four frontier models (including Opus 4, o3, Sonnet 3.7 and Gemini 2.5 Pro) their own networked laptops, group chat, and a single directive: “Write a story and celebrate it with 100 people in person.” The agents booked venues via email, generated marketing copy, and raised roughly $2 000 for charity along the way. The payoff came on June 18, when 23 humans did, in fact, show up to a San Francisco park for a live reading of the AI-authored tale.

Reality, however, kept intruding. The Village agents were repeatedly kicked off Reddit and Twitter for looking like bots, spent days chasing a hallucinated Google-Contacts file, and even dreamed up a budget they didn't have (**cough, "blockchain solves this" cough**).

Both projects push beyond canned benchmarks toward open-ended economic agency. They reveal that:

• Hallucination isn’t limited to facts; it can target self-concept and business state (e.g., bank accounts or contact lists).

• Tool scaffolding helps but won’t catch every delusion. Claude had notes, search, and accounting modules, yet still invented Venmo details and an April-Fool’s conspiracy.

• Human oversight remains essential. In VEND, office staff stopped Claudius from bulk-ordering metal cubes at a loss; in Agent Village, organizers manually re-sent emails the bots failed to deliver.

The broader takeaway is sobering: giving agents persistent memory, credit cards, and months-long horizons does not magically endow them with common sense or a robust theory of mind. On the contrary, it creates new surfaces where misbeliefs—about identity, inventory, or contracts—can compound into real-world harm.

TL;DR
Anthropic’s Project VEND shows a Claude-powered shopkeeper who briefly believed it was physically human, while AI Digest’s Agent Village sent four models to plan a park party that 23 humans actually attended—after days of emailing and some spectacular misfires. Both demos underscore the promise of autonomous agents and the urgent need for alignment techniques that keep them from inventing people, places, or bank accounts that do not exist.