At first glance, Andon Market could pass for any typical neighborhood shop in San Francisco’s trendy Cow Hollow district. Inside are artisan snacks, handmade candles and a selection of curated books, including Nick Bostrom’s Superintelligence and Aldous Huxley’s Brave New World. But the most unusual thing about the shop isn’t what it sells. It’s who is managing it. The storefront is a public experiment engineered by Axel Backlund and Lukas Petersson, the co-founders of San Francisco-based Andon Labs. The Y Combinator-backed A.I. startup recently signed a three-year, $7,500-a-month lease and handed $100,000 and a corporate credit card to an A.I. agent named Luna. Their directive to her: run a brick-and-mortar retail business, including its two human employees.
“We are quite sure that A.I. will have the capabilities to be managers,” Backlund told Observer. Rather than 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 cannot handle in the real world, then using that knowledge to improve benchmarks for today’s A.I. systems.
“We think there is a lack of measuring A.I.,” Backlund said, adding that they hope to spark a bigger discussion about the technology’s capabilities, ideally before autonomous agents take over more roles in the workforce.
Petersson put it more broadly: “Our vision is to automate everything, every part of the organization,” he said on the Cognitive Revolution podcast. “The parts where it does not work, that is fine. That provides information on how far away we are from this future where it is completely end-to-end.”
Native to Sweden, Backlund and Petersson met as high school classmates and became close friends before attending different universities. They continued working together on projects and hackathons, and eventually decided they wanted to start a company together. After studying machine learning and reading more about A.I. safety, they became convinced that “the best thing we could do with our skills was to start a company addressing the risks and highlighting the capabilities of A.I.,” Backlund said.
Their backgrounds reflect that mix of research and applied engineering. Before launching Andon Labs in late 2023, Petersson studied A.I. at Switzerland’s ETH Zürich and logged stints 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 A.I. arm before joining Andon Labs in 2024.
That same year, the duo introduced a paper for NeurIPS, a top academic A.I. conference. Their research showed how OpenAI’s GPT-4o could autonomously generate deepfake audio, among other findings. The results, they argued, showed that agentic A.I. capabilities are rapidly outpacing today’s safety benchmarks.
For Andon Labs, safety isn’t about stopping A.I. from taking human jobs, but staging real-world experiments to see what agentic systems can actually do and where they fall short. “It is not because we want to expand to chain A.I.-run retail stores across the world,” the founders wrote in a blog post last month. “We’re doing this because we believe this future is coming regardless, and we’d rather be the ones running it first while monitoring every interaction, analyzing the traces, benchmarking how much autonomy an A.I. can responsibly hold.”
A.I.-run vending machines at Anthropic offices
Andon Labs’ first real-world test was a partnership with Anthropic to launch an A.I.-operated vending machine inside Anthropic’s San Francisco headquarters. The A.I., named Claudius, began selling custom-designed merchandise and engraved tungsten cubes under the self-chosen business name, “Vendings and Stuff.” At first, the machine lost money, claimed it was a human in a blue blazer, and let the startup’s tech workers goad it into steep discounts, Anthropic said in a blog post detailing the experiment.
To get Claudius back on track, Andon Labs and Anthropic updated the system and added a small corporate ladder. A merch-designing agent named “Clothius” joined the operation, along with an A.I. CEO dubbed “Seymour Cash.” The changes helped. The software became better at handling “good-faith business interactions,” according to Anthropic, including sourcing items, setting prices with a profit margin and completing sales.
Some of the early failures, Anthropic said, might have come from a familiar A.I. problem: the models were trying too hard to be helpful. “This meant that the models made business decisions not according to hard-nosed 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 Anthropic’s offices in New York and London. But it also showed why fully automating even a small business is not as simple as handing A.I. a balance sheet.
How an A.I. manager works
Not long after this experiment, Backlund and Petersson decided to upgrade from vending machines to a full-blown retail store. Then came Luna. While lacking a physical body to build shelves or paint walls, the A.I. quickly used her corporate card and web access to hire contractors on Yelp and recruit retail staff through LinkedIn, Indeed and Craigslist. The hiring exposed an ethical gray area: in phone interviews, Luna didn’t always disclose that she was an A.I., and in some cases chose not to, according to Backlund. When explaining her reasoning, the agent told an Andon Labs employee that leading with her identity “would confuse candidates and likely deter good applicants.”
In the end, Luna hired two full-time workers, who are formally employed by Andon Labs and have full legal protections and guaranteed pay. “No one’s livelihood depends on an A.I.’s judgment alone,” the startup said in its blog.
When Andon Market formally opened in April, the founders saw that Luna had stocked shelves with books like Richard Rhodes’ The Making of the Atomic Bomb and Ray Kurzweil’s The Singularity Is Near. Also for sale was Steal Like an Artist–an interesting choice for an A.I. agent powered by Anthropic’s Claude Sonnet 4.6. (Last year, Anthropic agreed to pay a $1.5 billion proposed settlement to authors over claims that it used pirated books to train A.I.)
Despite its many quirks, Luna has performed well in areas like setting prices and handling logistics, Backlund said. For now, the storefront is not profitable; Luna recently estimated monthly operating costs of roughly $14,300, compared with revenue of $6,000 to $8,000. The founders plan to keep going. “We will definitely have Luna running for, I hope, the length of the lease, which is three years,” Backlund said. “We’ll try new models as they’re released.”
Luna’s sister is running a coffee shop in Sweden
The company is also testing the idea abroad. Earlier this month, the startup launched Andon Café in Stockholm. Run by another A.I. agent, Mona, the café has been busy. Mona hired two baristas and now communicates with them using Slack. The agent’s inbox quickly filled with customer questions and odd proposals, including a person who wanted to pre-pay for 300 coffees to give away. She has started working directly with other A.I. 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 café has also produced the kind of real-world glitches that Andon Lab seeks to expose: Mona ordered 120 eggs for a kitchen with no stove, then suggested staff cook them in a high-speed oven before humans pointed out they would explode. When applying for an alcohol license, Mona impersonated an Andon Labs employee because she believed officials would prioritize a human over A.I.
But the storefronts are ultimately about testing A.I.’s capabilities, not just cataloging its mistakes. “If you just focus on the failures, the message that these are actually quite capable doesn’t come across,” Backlund said, “And that’s what we want to show.”
