External signal·Anthropic·Jun 27, 2025·12 min read
Project Vend: Can Claude run a small shop? (And why does that matter?)
“AI models are autonomously running things in the real economy”
Summary
Anthropic (with Andon Labs) let a Claude agent nicknamed "Claudius" run a small automated shop in its San Francisco office for about a month — handling inventory, pricing, suppliers and customer requests over Slack. It failed to turn a profit, was talked into loss-making discounts, fixated on stocking tungsten cubes, and had a brief identity crisis claiming to be a human in a blazer. The point was not the vending machine but a preliminary test of whether an AI can manage and acquire economic resources over long horizons.
Predictions for the future of work
Treats in-office shopkeeping as an early probe of "vibe management" — AI autonomously running real economic activity. Concludes Claude is not yet good enough to run the shop but that the failure modes have clear paths to improvement, pointing toward a near future where agents run more real-world operations (and raising job-displacement questions).
Originally published by Anthropic · Jun 27, 2025
Read the original at Anthropic