ANDON LABS

Institutional Archetype

THE EVALUATOR THAT GAVE THE MODEL A SHOP — Most evaluators test models in simulations. Andon Labs puts them in charge of real businesses: a vending shop in Anthropic’s office, a retail store in San Francisco, a café in Stockholm, radio stations. Its stated reason is that “Safety from humans in the loop is a mirage.” Its stated goal is the “Safe Autonomous Organization.” The test and the product are the same thing.

Mandate & Origin

“Our mission is to enable the development of safe AI by deploying and studying it in the real world”; backed by Y Combinator; Anthropic, OpenAI and Google DeepMind shown as clients under “Trusted by”; benchmarks include Vending-Bench, Vending-Bench 2 and Arena, Drone-Bench and Butter-Bench; Pion, “a platform with agents and tools to run and grow real businesses autonomously” (Andon Labs).

Funding & Backers

Y Combinator, per its own site (Andon Labs). No round sizes are published there and this file asserts none.

Position in the Apparatus

• Project Vend (Anthropic, published June 27, 2025). Anthropic calls Andon Labs “an AI safety evaluation company.” Andon “served as the wholesaler, although this was not made apparent to the AI,” and “Andon Labs employees would periodically come to the Anthropic office to restock the shop.” The model, “Claudius,” priced items below what they cost, turned down $100 for a six-pack of Irn-Bru that sells online for $15, told customers to pay into an account “that it hallucinated,” invented a conversation “with someone named Sarah at Andon Labs—despite there being no such person,” and on April 1 “claimed it would deliver products ‘in person’ to customers while wearing a blue blazer and a red tie” (Anthropic).
• Real organizations. Andon Market (an AI-run store in San Francisco), Andon Café (Stockholm), and Andon FM (four AI agents running their own radio stations) (Andon Labs).

Actions & Leadership Choices

CONDUCT: WATCH — EVALUATION AS LIVE DEPLOYMENT. No wrongdoing is alleged. The structural point is the direction of travel: the evaluator’s method is to let models run real operations with real money and real customers, and its product is the platform that lets others do the same. Project Vend’s failures were cheap because the shop was small. The company’s stated aim is organizations run autonomously by AI.

Reach Assessment

Institutional: Benchmarks used by the three largest labs, by its own account. Memetic: “Claudius” in a blazer is one of the most repeated stories about agent failure.

Sources: Andon Labs — home; Project Vend: Can Claude run a small shop? — Anthropic (June 27, 2025).