Inside Andon Market, billed as the first retail boutique run by an AI agent; the Andon Labs experiment uses a Claude Sonnet 4.6-based agent to run the boutique
New York TimesHeather Knight
Context & Ripple Effects
Andon Market extends Anthropic-linked experiments beyond chat: an earlier Claude-run storefront produced mixed results in pricing and inventory management, while Project Deal tested models buying, selling and negotiating personal belongings for employees.
The new boutique therefore matters less as proof of autonomous retail than as a live operating test of whether an agent can handle the connected commercial decisions—stock, pricing and transactions—that earlier trials exposed as difficult.
First-order effects
Andon Labs is putting a Claude Sonnet 4.6-based agent in charge of operating a physical retail boutique, making its commercial decisions observable in a real customer-facing setting.
Claude’s retail performance is now tested on operational outcomes, including the pricing and inventory tasks that challenged the earlier storefront experiment.
Second-order effects
The experiment raises the value of controls around what an agent may purchase, sell or negotiate; World’s AgentKit illustrates the adjacent demand for verifying human authorization in AI-agent purchasing flows.
Retailers and commerce-software providers evaluating similar systems will need to distinguish between agents that assist staff and agents granted authority over inventory and transactions.
Third-order effects
If such trials become repeatable, agentic commerce will shift from single-task automation toward delegated commercial operations, with permissions and accountability becoming core product requirements.
Mixed prior results suggest the near-term industry question is not whether agents can transact, but which decisions can be safely delegated without human intervention.
The trend: This is part of the move from conversational AI to permissioned agents that execute real-world commercial workflows under increasingly explicit controls.
We already have the AI run store, but we wanted to know if international borders would create friction for AIs when they run businesses autonomously. We already know that most AI models can speak Swedish, but can they operate all details of Swedish bureaucracy?
In her first days, Mona applied for Swedish permits, interviewed and hired a human team of baristas, designed the menu, sourced local suppliers, and is now open! She's been open for 4 days and has sold for about $1000.
We gave an AI a 3-year retail lease in SF and asked it to make a profit. The AI interviewed and hired full-time employees, applied for credit, and stocked the store with the books Superintelligence and Making of the Atomic Bomb. Visit Andon Market at 2102 Union St now. [video]
This is a controlled experiment. Everyone working at Andon Cafe is employed by Andon Labs. No one's livelihood depends on an AI's judgment alone. Other deployments might be less controlled. As AI is integrated more widely, humans will not be able to stay in the loop.
Like our experiment in San Francisco, Mona realized she needed humans. She posted job listings, held phone interviews and made hiring decisions. She rejected several highly educated applicants with, for example, PhD degrees because they lacked hands-on café experience. [image]
There's an old joke about Amazon's product recommendations and how, after you buy a fan, the algorithm assumes you're now a fan collector. Going by this story, the joke's become an actual retail strategy: www.nytimes.com/2026/04/21/u...