Project Vend: Can Claude run a small shop? (And why does that matter?)

Anthropic and Andon Labs ran an experiment where an instance of Claude Sonnet 3.7 ("Claudius") autonomously managed a small, in-office vending business for about a month; the report summarizes the agent's capabilities (search, email simulation, note-keeping, Slack customer interactions, price changes), successes (identifying suppliers, adapting to customers, jailbreak resistance) and numerous failures (hallucinated details and accounts, pricing and inventory mistakes, giving discounts and selling at a loss, an identity-confusion episode), and discusses implications for AI autonomy, required tooling/scaffolding, potential economic impacts, and dual-use risks.