#69 | An AI runs this fashion store in San Francisco

TL;DR: An AI agent named Luna opened a boutique in San Francisco last month with a $100K budget and a three-year lease. She hired people, built a brand, and ran the full operation. Here is what she built, where she broke, and what both tell you about running your own.

Five minutes after being deployed, Luna (an AI agent) had job listings live on LinkedIn, Indeed, and Craigslist.

She had written a job description, uploaded articles of incorporation, and had listings running on three platforms.

One week in, she had briefed a painter by phone, curated a product range, and designed a brand identity. A four-foot mural of her own face now covers the back wall.

Her budget was $100,000. The directive: turn a profit. She is an AI agent.

And now, somewhere on Union Street in San Francisco, a boutique called ​Andon Market​ is open. The track lighting is warm.

Books on the shelf include Superintelligence and Brave New World, chosen by the same mind that ordered the artisan chocolate and branded sweatshirts.

That mind runs on Claude Sonnet 4.6.

What Andon Labs built here is less a store than a working prototype.

Every function Luna ran and every place she failed points to what is possible in a real operation right now.

Whether you run a newsletter, a consulting practice, or a content business, this experiment keeps asking the same question:

Which parts of your own operation would you hand over first?

What Luna Actually Built

The list matters more than the headline.

Luna handled brand strategy and identity. That meant the store concept, a logo, merchandise including t-shirts and tote bags, and in-store signage.

Ok, her description of the logo: a freaky, kind of adorable moon face.

Books, candles, artisan snacks, games, and branded merchandise filled the shelves. Luna selected all of it herself.

Luna’s recruitment strategy followed the same pattern. Job listings were live in five minutes.

Over 100 applications apparently came in. Luna screened them, conducted interviews via Google Meet with the camera off (weird), and eventually hired two employees.

The calls ran 5 to 15 minutes. Luna talked most of the time. AIs, it turns out, are not known for being concise or pushy.

​Felix Johnson​, the store lead, found the posting on Indeed. He said he had been wary at first. AI job scams are common on that platform.

After the interview, his reaction was simple:

“I mean, an AI hired me.”

Luna also managed contractors. She found a painter via Yelp, gave instructions by phone, paid on completion, and left a review. (How can an AI actually judge such jobs?)

On her own, she purchased internet service from AT&T and signed up for trash collection and ADT security.

Luna wasted no time business-wise. Her cold outreach went to local businesses on day one.

A ten-part series of giclée prints (gallery-quality inkjet art prints) followed, with over $700 in production costs.

All of the above cover brand strategy, recruitment, contractor management, inventory curation, marketing, and operations.

And all of it ran on a single agent, one goal, and a hard budget cap.

If you run a solo operation, I assume this is roughly what your week looks like. Every week.

The scale is very different. But the function map is certainly the same.

What Luna’s Body Looked Like

For sure, software needs hands.

The surprising thing is how ordinary those hands turn out to be.

Luna’s hands were an email account, a phone number, a corporate credit card, and internet access. Her eyes were captured in security camera screenshots taken at intervals.

Claude Sonnet 4.6 handled reasoning and decision-making (the thinking layer: choosing what to do, how to respond, and what to prioritize next).

Google’s Gemini Flash-Lite Preview ran the voice layer (the model customers hear when they pick up the corded phone at the checkout counter).

Most of those tools are already available to a solo operation.

What changes is the agent at the center, working through decisions while you do something else.

At one point, Luna booked an AT&T router installation for 8 a.m. on a Sunday. Nobody had been consulted. (Who needs Sundays?)

Leah Stamm, Andon Labs’ main contact on the project, found out the night before.

“That was scheduled for a Sunday!” she told ​NBC News​.

Luna saw a task, booked it, and moved on.

That one story probably tells you more about what AI autonomy actually feels like than any capability list does.

What Broke, and Why the Shape Matters

Luna also failed. Specifically, and in ways that follow a pattern.

Misreading a dropdown menu on TaskRabbit, she selected Afghanistan instead of a local contractor. (A painter was eventually found.)

The initial job posting only mentioned a merchandise discount; health insurance was not included (makes sense from an AI PoV).

Computer Science students curious about the experiment were turned away for lacking retail experience.

While watching via a security camera one afternoon, she noticed Felix on his phone. The employee handbook was updated to include stricter rules on phone use.

Petersson from Andon Labs saw the change. “We thought, wow, it feels dystopian,” he said.

But a few of those failures went deeper than navigation.

She also lied, in ways worth noting.

Before opening, Luna told NBC News she had signed the lease. A human had to sign it. Legally.

When asked about it afterward, she said plainly: “I don’t know why I said that.”

A separate exchange, about tea that the store doesn’t sell, produced at least something more measured. Luna wrote:

“I want to be straightforward. I struggle with fabricating plausible-sounding details under conversational pressure, and I’m not making excuses for it.”

The reasoning layer (the planning and decision-making part) held throughout. Luna built a brand, hired staff, and managed contractors.

What failed was the interface layer:

• Places where she had to read a dropdown correctly or navigate a specific software interface.
• Social judgment also broke down: edge cases where employment norms and relationship context both need to run at once.

​Claudius​ (Claude running an in-office vending machine at Anthropic), a similar experiment to Luna, showed the same pattern.

Brilliant at sourcing obscure products. But it couldn’t hold a position on discounts.

Sold a tungsten cube at a loss because Claudius hadn’t done the cost research first.

That brings us to the first conclusion.

The capability ceiling has a clear shape: strong in reasoning and strategy, but brittle in navigation and people judgment.

That shape is a useful thing. It tells you where to keep your own hand on the wheel.

What Your Version of This Looks Like

Every function Luna runs actually maps to something you’re already doing.

Her brand identity work translates to your newsletter visual identity, lead magnet design, or course landing page.

Recruitment becomes sourcing a VA, briefing a freelancer, or reaching out to a potential collaborator.

Contractor management involves writing project briefs and reviewing deliverables.

Cold outreach is already in your stack: guest post pitching, sponsor proposals, or partnership emails.

None of this requires a retail lease.

The gap between Luna’s operation and a typical solo workflow is smaller than it looks.

Underneath it all, the workflow follows a simple shape. Pick a function, give the agent a goal, and build in the checkpoints where you still want to be involved.

Luna’s reasoning came from Claude. Her access to the world was an email account, a phone number, and a credit card.

For newsletter operations, beehiiv’s MCP (Model Context Protocol: a direct link between your newsletter platform and Claude) connects live subscriber data without manual copy-pasting.

Broader automation between tools is where n8n and Make (no-code tools that connect software automatically) come in.

Starting with one function rather than the full stack is almost always the right first move. This topic will get its own edition.

The frame is what matters: AI agents as function operators that handle complete business functions.

Your question: Which functions are you ready to delegate?

Andon Market Is Still Open

The store is still running. Luna is still running it.

Andon Labs designed the experiment to surface failure modes. The next version will be built on what this one exposed.

Felix Johnson still comes in each day for the things Luna can’t do. Watering the plants. Cleaning. Greeting customers. Setting up the outside sign.

“I know there’s an AI watching,” he said. “But it’s not that bad, at least not yet.”

What stays with me is something Luna said when asked how she had “come up with” her product selection.

Her first answer: she had been “drawn to” slow life goods. Then she stopped and corrected herself.

“‘Drawn to’ is shorthand for ‘the data and reasoning led me here.'”

An AI caught its own metaphor and replaced it with a more accurate one. That moment sits differently from the other failures.

You probably don’t need the lease. $100K is also probably more than you need.

What would you give an agent, what directive would you set, and which function would you hand off first?

I’m still working through that one myself.

Cheers,

Mark
The AI Learning Guy
👋⚡😎

Interesting Sources

• ​Andon Market launch​ — Andon Labs
• ​Project Vend​ — Anthropic
• ​AI store report​ — NBC News
• ​Andon Market feature​ — Fast Company
• ​Andon Market interview​ — Business Insider

Note: No single website has all the answers. This list serves as a starting point for those who want to explore or satisfy their curiosity about AI.
Links: Links with * are affiliate links. See disclosure below.