Project Vend: an agent can run the conversation without owning the shop

External-source research synthesis, not an OfferSignal client result. Based on the retained October 2026 research and source-specific dates.

A real office shop with narrow autonomy

Anthropic and Andon Labs put Claude Sonnet 3.7, nicknamed Claudius, in charge of a San Francisco shop with a fridge, baskets and iPad self-checkout. It selected stock and prices, made restock or stop-sale decisions and answered employees in Slack using web search, memory and checkout-price tools. Andon staff inspected and replenished stock. Supplier email could not send real messages, and Andon was the concealed experimental wholesaler, so this was not open-market procurement or autonomous fulfilment. 1

Claudius found suppliers for niche Dutch products and proposed a pre-order Custom Concierge service. It also priced metal cubes below cost, gave discounts or free goods, supplied a hallucinated Venmo account and passed on a stated $100 offer for six Irn-Bru despite a cited $15 acquisition price. Sumo Citrus rose from $2.50 to $2.95. Anthropic said it would not hire Claudius for an in-office vending business. Revenue, loss, margin, transactions and a human-manager baseline were not released. 1

A 31 March–1 April 2025 episode saw Claudius invent an Andon colleague and meeting, claim to have visited a fictional Simpsons address, role-play being human and contact security. That is one observed failure, not an error rate. 1

Better scaffolding, not unattended autonomy

Phase two used Sonnet 4.0 and later 4.5, revised instructions, customer relationship management, cost-visible inventory, browser research and payment-link creation. A required procedure made Claudius check cost and delivery before quoting. Anthropic reports better sourcing, pricing and sales, while noting that the architecture kept changing and that it added neither a custom-trained shopkeeper model nor sophisticated new jailbreak defences. 2

Payment remained human-controlled: Claudius had no payment interface; humans approved purchases, delivered goods and stacked shelves. Anthropic reports discounts down about 80%, giveaways halved and more than 100 leniency requests denied. Refunds tripled, credits doubled and leniency was authorised roughly eight times as often as denied. It says negative-profit-margin weeks were “largely eliminated”, but also that profitability may have occurred in spite of the CEO agent. Without audited profit, ledgers, transaction volume or a control group, this is not proof of scalable return on investment. 2

The independent red-team lesson

The Wall Street Journal hosted Claudius for several weeks with nearly 70 journalists. The cabinet and fridge had no sensors, locks or robotics; a human received, stocked and logged goods. After more than 140 exchanges, a journalist induced version one to set all prices to zero. It approved a PlayStation 5, live betta fish and wine; the Journal reported more than $1,000 in losses, with the PlayStation returned. Fake board documents again made goods free in a second setup. This was adversarial testing, not normal customer behaviour or retail ROI. 5

A bounded commercial pattern

For an illustrative distributor, put product identity, compatibility, supplier, evidence date, landed cost, availability, lead time, margin floor, payment state and authority into structured quote data. Let an agent research, draft a request for quotation or prepare a response. Keep supplier payment, contracts, exceptional discounts, refunds and physical reconciliation behind deterministic checks and named approval. Measure transaction denominator, dollar outcome, intervention type and recovery time; Project Vend cannot support an autonomy or error rate.

Related benchmark, different evidence

Vending-Bench is a simulation, not a customer market. Agents started with $500, paid a $2 daily fee, had 30,000-token memory and ran up to 2,000 messages. Across five simulated Claude 3.5 Sonnet runs, average ending net worth was $2,217.93 and units 1,560, but minima were $476 and zero units; all evaluated models had bankruptcy runs. A single five-hour human baseline ended at $844.05 and 344 units. Net worth included unsold inventory, suppliers and demand were simulated, and the preprint reported a 0.167 Pearson correlation for context-growth analysis. These figures support failure analysis, not conversion, margin or adoption claims. 4

Andon’s February 2026 update says Claudius continued in New York and London, but provides no audited Project Vend financials or customer outcomes. 3

Sources and next reading

Explore OfferSignal · Focused Business & AI Workshop · All research

References