At first glance, Andon Market could be considered any typical store in San Francisco’s trendy Cow Hollow neighborhood. Inside are handcrafted snacks, handcrafted candles, and a selection of curated books, including those by Nick Bostrom Super intelligence And Aldous Huxley Brave new world. But the most unusual thing about the store isn’t what it sells. He’s the one who runs it. The storefront is a public experience designed by Axel Backlund and Lukas Peterson, co-founders of San Francisco-based Andon Labs. The Y Combinator-backed AI startup recently signed a three-year lease for $7,500 per month and handed over $100,000 and a company credit card to an AI agent named Luna. Their guidance has: retail business management, incl its human employees.
“We are very confident that AI will have the capabilities to become a manager,” Backlund told the Observer. Instead of chasing retail margins, Backlund and Petersson are using the newly launched storefront to stress-test what Luna, powered by Anthropic’s Claude model, can and can’t handle in the real world, then use that knowledge to improve benchmarks for existing AI systems.
“We think there is an under-measurement of AI,” Backlund said, adding that they hope to spark a larger discussion about the capabilities of the technology, ideally as autonomous agents take on more roles in the workforce.
Peterson put it more broadly: “Our vision is to automate everythingHe said in every part of the organization The cognitive revolution Podcast. “The parts where it doesn’t work are okay. This provides information about how far we are from this future where it’s completely universal.”
Backlund and Peterson, a native of Sweden, met as high school classmates and became close 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 reading more about AI safety, they became convinced that “the best thing we can do with our skills is to create a company that addresses the risks and highlights the capabilities of AI.” Backlund said.
Their backgrounds reflect this combination of applied research and engineering. Before launching Andon Labs in late 2023, Peterson studied artificial intelligence at ETH Zurich in Switzerland and logged stints at Google and Disney Research. Backlund, who studied at KTH Royal Institute of Technology in Sweden and at EPFL in Switzerland, worked as a data engineer at McKinsey’s artificial intelligence arm before joining Andon Labs in 2024.
In the same year, the duo presented a paper to NeurIPS, a high-level academic conference on artificial intelligence. Their research showed how OpenAI’s GPT-4o can autonomously generate deep fake audio, among other results. They said the results showed that agent AI capabilities are rapidly outperforming current safety standards.
For Andon Labs, safety is not about preventing AI from taking over human jobs, but rather about conducting real-world experiments to see what agent systems can actually do and where they fall short. “Not because we want to expand into an AI-run retail chain “We do this because we believe this future is coming regardless, and we would rather be the ones to manage it first as we monitor every interaction, analyze the implications, and measure how much autonomy AI can responsibly have,” the founders wrote in a blog post last month.
AI-operated vending machines in Anthropy’s offices
Andon Labs’ first real-world test was partnering with Anthropic to launch an AI-powered vending machine inside Anthropic’s headquarters in San Francisco. The artificial intelligence, named Claudius, began selling specially designed merchandise and engraved tungsten cubes under his self-chosen brand name, “Vendings and Stuff.” At first, the machine lost money, pretended to be a human in a blue jacket, and allowed the startup’s tech workers to push it into deep discounts, Anthropic said in a blog post. Detail the experience.
To get Claudius back on track, Andon Labs and Anthropic updated the system and added a small ladder to the company. A merchandise design agent named “Clothius” is joined in the operation, along with an AI executive named “Seymour Cash”. The changes helped. The software has gotten better at handling “good faith business interactions,” according to Anthropic, including sourcing items, setting mark-up prices, and completing sales.
Anthropic said some of the early failures may have come from a familiar problem with AI: models were trying too hard to be useful. “This means that the models made business decisions not according to strict market principles, but from something more like the perspective of a friend who just wants to be nice.”
The vending machine business eventually became profitable and expanded to include Anthropic’s offices in New York and London. But it also showed why full automation of even small businesses isn’t as simple as handing over the balance sheet to AI.
How does an artificial intelligence manager work?
Shortly after this experience, Backlund and Peterson decided to upgrade from vending machines to a full-fledged retail store. Then Luna came. Despite lacking a physical body to build shelves or paint walls, the AI quickly used its company card and web access to hire contractors on Yelp and recruit retail employees through LinkedIn, Indeed, and Craigslist. Hiring revealed an ethical gray area: In phone interviews, Luna did not always disclose that she worked in AI, and in some cases she chose not to, according to Backlund. When explaining her reasons, the agent told an Andon Labs employee that using her identity “would confuse candidates and potentially deter good applicants.”
Eventually, Luna hired two full-time workers, who were officially employed by Andon Labs and enjoyed full legal protection and guaranteed wages. “No one’s livelihood depends on the judgment of AI alone,” the startup said in its blog.
When Andon Market officially opened In AprilThe founders saw that Luna had filled the shelves with books like those of Richard Rhodes Making the atomic bomb And Ray Kurzweil The singularity is near. It was also for sale Steal like an artist– Interesting choice of AI agent powered by Anthropic’s Claude Sonnet 4.6. (Last year, Anthropic agreed to pay A Proposed settlement worth $1.5 billion to the authors over allegations that she used pirated books to train artificial intelligence)
Despite its many quirks, Luna has performed well in areas such as setting prices and handling logistics, Backlund said. Right now, the storefront is not profitable; Luna Recently estimated Monthly operating costs are approximately $14,300. This compares to revenues of between $6,000 and $8,000. The founders plan to continue. “Luna will certainly be there, I hope, for the duration of the lease, which is three years,” Backlund said. “We will try new models as they are released.”
Luna’s sister runs a café in Sweden
The company is also testing the idea abroad. Earlier this month, the startup launched Andon Cafe In Stockholm. Run by another AI agent, Mona, the café was crowded. Mona hired two baristas and now communicates with them using Slack. The agent’s inbox quickly filled with customer questions and outlandish suggestions, including one person who wanted to prepay for 300 coffee cups to donate. I started working directly with other AI clients, and had a Google Meet meeting with someone who wanted to learn how to run a business.
In its first week, Andon Café generated sales of 10,000 Swedish krona, or about $1,068, according to Backlund. The café also produced the kind of real-life mistakes that Andon Lab seeks to expose: Mona ordered 120 eggs for a stoveless kitchen, then suggested the staff cook them in a high-speed oven before the humans signaled they would explode. When applying for an alcohol license, Mona impersonated an Andon Labs employee because she thought officials would prioritize humans over artificial intelligence.
But storefronts are ultimately about testing AI’s capabilities, not just cataloging its errors. “If you just focus on the failures, you’re not going to get the message that these devices are fully capable of doing this. And that’s what we want to show,” Backlund said.
