At first glance, Andon Market could pass for any nearby store in San Francisco’s trendy Cow Hollow neighborhood. Inside are artisan snacks, handmade candles and a wide selection of books, including Nick Bostrom’s. Superintelligence and Aldous Huxley Brave New World. But the most unusual thing about the store is not what we sell. He is the one who owns it. The storefront is a public experiment by developers Axel Backlund and Lukas Petersson, founders of San Francisco-based Andon Labs. A Y Combinator-backed AI startup recently signed a three-year, $7,500-a-month lease and gave $100,000 and a business credit card to an AI agent named Luna. Their instruction to him: run a brick and mortar retail business, incl its two human employees.
“We are confident that AI will have the skills to become managers,” Backlund told the Observer. Rather than chasing store margins, Backlund and Petersson are using the newly launched storefront to stress-test what Luna, powered by Claude’s Anthropic model, can and can’t handle in the real world, and then use that knowledge to improve the ratings of today’s AI systems.
“We think there’s a lack of scalable AI,” Backlund said, adding that they hope to spark a larger conversation about technical capabilities, before autonomous agents take over more roles in the workplace.
Petersson put it more broadly: “Our vision is to automate everything, every part of the organization,” he said in a statement. The Cognitive Revolution a podcast. “The parts where it doesn’t work, that’s fine. That gives some insight into how far we are from this future where it’s completely finished.”
A native of Sweden, Backlund and Petersson met as high school classmates and became best friends before attending different universities. They continued to work together on projects and hackathons, and eventually decided they wanted to start a company together. After studying machine learning and learning more about AI safety, they were convinced that “the best thing we could do with our skills was to start a company that deals with risks and highlights the capabilities of AI,” Backlund said.
Their backgrounds reflect that mix of research and applied engineering. Before launching Andon Labs in late 2023, Petersson studied AI at Switzerland’s ETH Zürich and interned at Google and Disney Research. Backlund, who studied at Sweden’s KTH Royal Institute of Technology and Switzerland’s EPFL, worked as a data engineer at McKinsey’s AI arm before joining Andon Labs in 2024.
That same year, the duo presented a paper at NeurIPS, a top AI conference for education. Their research showed how OpenAI’s GPT-4o can automatically generate deep audio, among other findings. The results, they argued, showed that the agent’s AI capabilities are rapidly surpassing today’s security benchmarks.
For Andon Labs, security is not about stopping AI from taking over human jobs, but about putting real-world tests to see what agent programs can do and where they fail. “It’s not that we want to expand to include AI-run stores around the world,” the founders wrote in a blog post last month. “We’re doing this because we believe this future is coming anyway, and we’d better be the first to drive it while monitoring every interaction, analyzing the trajectory, looking at how much autonomy AI can responsibly handle.”
AI driven sales machines in Anthropic offices
Andon Labs’ first real-world experiment was a collaboration with Anthropic to launch an AI-powered vending machine inside Anthropic’s headquarters in San Francisco. The AI, named Claudius, began selling custom-designed merchandise and engraved tungsten cubes under the business name of his choice, “Vendings and Stuff.” At first, the machine lost money, said the man in the blue blazer, and allowed the startup’s technical staff to inflate it, Anthropic said in a blog post detailing the research.
To get Claudius back on track, Andon Labs and Anthropic updated the system and added a small business ladder. A sales design agent named “Clothius” joined the project, along with an AI CEO named “Seymour Cash.” Changes helped. The software became better at managing “good faith business interactions,” according to Anthropic, including purchasing goods, setting prices for profit margins and completing sales.
Some early failures, Anthropic says, may have stemmed from a general AI problem: the models were trying too hard to be helpful. “This means that models make business decisions not based on strict market principles, but from something like the opinion of a friend who just wants to be nice.”
The vending machine business eventually became profitable and expanded to Anthropic offices in New York and London. But it also showed why automating even a small business isn’t as simple as giving AI a balance sheet.
How the AI manager works
Shortly after this test, Backlund and Petersson decided to upgrade from vending machines to a full-service retail store. Then Luna arrived. Although it doesn’t have a physical body to build shelves or paint walls, AI quickly used its business card and web access to hire contractors on Yelp and hire sales staff through LinkedIn, Indeed and Craigslist. The hiring exposed a gray area of behavior: in phone interviews, Luna did not always reveal that she was an AI, and in some cases she chose not to attend, according to Backlund. In explaining his thinking, the agent told an Andon Labs employee that leading by identity would “confuse candidates and deter good applicants.”
In the end, Luna hired two full-time employees, legally employed by Andon Labs and with full legal protections and guaranteed pay. “No human livelihood depends solely on the judgment of AI,” the startup said on its blog.
When Andon Market is officially opened in Aprilthe founders realized that Luna had installed shelves with books such as Richard Rhodes. The Making of the Atomic Bomb and Ray Kurzweil Unity Is Near. And it was for sale Be like a musician-an interesting AI agent option powered by Anthropic’s Claude Sonnet 4.6. (Last year, Anthropic agreed to pay ia $1.5 billion is proposed to be paid to the authors for claims that it used crime-ridden textbooks to train the AI)
Despite its many problems, Luna has performed well in areas such as pricing and handling, Backlund said. Currently, the storefront is unprofitable; Luna recently estimated monthly operating expenses of about $14,300, compared to an income of $6,000 to $8,000. The founders plan to continue. “We will definitely have Luna running for, I hope, the length of the lease, which is three years,” Backlund said. “We will try new models as they are released.”
Luna’s sister owns a coffee shop in Sweden
The company is also testing this idea abroad. Earlier this month, implementation began Andon Café in Stockholm. Run by another AI agent, Mona, the cafe was busy. Mona has hired two baristas and now communicates with them via Slack. The agent’s inbox was quickly filled with customer inquiries and strange proposals, including one who wanted to pay in advance for 300 coffees to brew. He has started working directly with other AI agents, holding a Google Meet with one who wanted to learn about running a business.
In its first week, Andon Café brought in 10,000 Swedish kronor, or about $1,068, in sales, according to Backlund. The restaurant also produced the kind of real-world problems Andon Lab wants to expose: Mona ordered 120 eggs from a kitchen without a stove, then suggested the staff cook them in a high-speed oven before people pointed out that they would explode. When applying for a liquor license, Mona pretended to be an employee of Andon Labs because she believed that officials would put people first over AI.
But the storefront is about testing AI’s capabilities, not just pointing out its flaws. “If you only focus on failure, the message that these are really capable doesn’t get through,” Backlund said, “And that’s what we want to show.”
