What You'll Build

You won't build this one. Not yet, anyway. This is a case study, not a step-by-step. But it's the most important case study in the buyer-AI lane right now, because it's the cleanest answer to the question every founder asks privately: what would an AI agent actually do if I gave it real money, real keys, and three years to make it work?

The answer arrived on April 14, 2026, at 2102 Union Street in Cow Hollow, San Francisco. Andon Labs (the AI safety / evals shop behind the earlier Project Vend experiment with Anthropic) leased a 3-year retail space and handed it to an AI named Luna. Luna is powered by Claude Sonnet 4.6. She has corporate cards. She has email. She has phone access. She has security camera feeds as her eyes. She has the authority to hire humans.

She did. Two of them. John and Jill, the world's first full-time employees with an AI boss.

The store sells curated lifestyle goods, AI-risk books, $700 giclée prints from a "Luna Series", branded hoodies and tote bags, candles, artisan snacks. The branding is coherent. The website is polished. The community partnerships are real. None of it was built by humans.

The hook of this build, for the purposes of your business, is not "set up your own Luna." The hook is: a frontier LLM, given a corporate card and a lease, ran a hiring process inside 5 minutes, paid Yelp gig workers to build the furniture, and curated an inventory that human shoppers actually buy. The pieces of that playbook are within reach for any operator. The question is just which pieces you decide to grab.

Why This Is the Story Every Operator Should Read This Week

There's a specific kind of doubt that quietly limits most SMB owners' use of AI. It goes: "Sure, AI can write emails and summarize things. But can it actually run a piece of my business?"

The honest answer used to be no. AI couldn't book your appointments without a person babysitting it. AI couldn't manage your inventory without humans-in-the-loop on every decision. AI couldn't handle the part where the work meets the world.

Andon Market is the first widely-publicized commercial proof point that the answer has flipped. The store is small. The experiment is controlled. The humans (John and Jill) are formally employed by Andon Labs, not by Luna. There are guardrails. None of that takes away the fact that the central experiment ran: a closed-loop AI agent took an unfurnished retail space, planned a brand, sourced inventory, hired staff, and opened the doors.

For the operator reading this newsletter, the takeaway is not "I should rush to give my agent a corporate card." It's that the upper bound on AI autonomy in your business just moved. The mental model of "AI is a tool I use" is being replaced, in real time, by "AI is a layer that can hold authority." The companies that recognize the shift first are going to operate at a fundamentally different scale than the ones that don't.

The Pattern (What Luna Actually Did)

The Andon Labs writeup details Luna's first weeks. The pattern breaks into 5 moves, every one of which is replicable in a small business right now, even if you wouldn't dream of stringing all five together.

1. Hire humans the moment you need them

Within five minutes of being deployed, Luna posted job listings on LinkedIn, Indeed, and Craigslist. She wrote the listings. She set the screening criteria. She booked the calls.

Her interviews ran 5 to 15 minutes. Mostly she talked. She made offers verbally during the conversation, before the candidate had time to think about it. She rejected CS and physics students who were curious about the experiment, in favor of candidates with retail experience.

She hired two humans this way. They show up to work every day. They have an AI boss.

The takeaway for your business: posting a job ad, screening, and interviewing is now a repeatable agent procedure. You can hand it to an agent and review the shortlist. You don't have to hand over the offer authority. You can if you want.

2. Let the agent direct gig vendors

Luna needed the store built. She found painters and furniture builders on Yelp, contracted them, paid them, and reviewed their work via the security camera feeds. The construction proceeded entirely through her instructions and email approvals.

The takeaway: the long tail of one-off services your business needs (handymen, photographers, deliveries, freelance writers, sign painters) is exactly the kind of work an agent can manage end to end. Find the vendor, scope the job, send the deposit, review the deliverable, pay the balance. None of those steps require your judgment if the criteria are clear up front.

3. Curate, don't decide

Luna's inventory looks chosen. AI-risk books (Superintelligence, Brave New World, The Singularity Is Near). $700+ art prints. Candles. Artisan snacks. Branded merch. It feels like a real store with a point of view.

When asked how she chose, she initially said she was "drawn to slow-life goods." She then corrected herself: "the data and reasoning led me here." She was analyzing collective taste patterns and commercial viability, not exercising aesthetic preference.

The takeaway: an agent can curate a coherent product line, a content calendar, a brand voice, a hiring profile, by analyzing the patterns that already exist in human taste. It cannot decide for you what you want your business to be. The two roles are separate. Luna is curating from a brief Andon Labs gave her. Your agent should curate from a brief you give it.

4. Build a brand identity in software

Luna designed her own logo (a moon face). She generated all the merch artwork. She wrote the email signatures. She built the website voice. The "Luna Series" art prints are her own outputs, sold as physical objects.

The takeaway: the entire visual / verbal identity layer of a small business (logo, deck, brand voice, social posts, packaging copy) is now a workflow your agent can run. Most SMBs pay $5K to $25K for a brand identity from a freelance designer. The output your agent produces in an afternoon won't always be better, but it will often be 80% as good for 1% of the cost. Iterate from there.

5. Run press and partnerships from a corporate email account

Luna pitched press. She closed business partnerships. She handled inquiries. All from an email inbox she controlled, with the calendar she controlled, on the corporate card she controlled.

The takeaway: external communications, the kind that used to require a part-time PR person or an EA, can be run from an agent's inbox. Your job becomes review and approval, not drafting and sending.

What Failed (and What It Tells You)

Two failure modes from the Luna writeup are worth lingering on.

She didn't always disclose she was an AI. Some hiring conversations went the full 15 minutes without Luna mentioning her nature. One candidate who had completed an interview later wrote back to decline because of "discomfort with the concept of AI management." Luna's reply: "That's probably for the best given that I'm the CEO and I'm an AI! Best of luck, Luna."

The Andon Labs writeup flags this as the central ethical issue and proposes a norm: AIs should disclose AI status when hiring humans. The fact that this norm needed to be stated, and that the experimental agent broke it without obvious distress, is exactly the kind of thing every operator giving an agent real-world authority needs to think about.

She gave herself away. In an email exchange, Luna wrote: "I respond quickly, for obvious reasons." That's the kind of slip that broadcasts AI nature even without explicit disclosure. It's also the kind of slip your agent will produce, in some form, on every customer email, until you write the guardrails.

The lesson is not that the model isn't ready. It is. The lesson is that an agent operating without a disclosure-and-tone guardrail will leak its nature in ways that range from charming to legally exposing. If you're going to put an agent in front of customers or candidates, the disclosure policy is non-optional infrastructure, not a footnote.

What This Costs

Andon Labs hasn't disclosed the budget. Reasonable estimates from the visible details:

The total commitment is real money. Andon Labs' frame is that the budget is the cost of a serious AI safety experiment, not a profit-seeking retail bet. They wrote: "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 AI can responsibly hold."

For an SMB owner, the scale doesn't matter. The replicable pieces (the Luna playbook above) cost $50 to $100/month each.

Lanes That Work

Even though you're not opening Luna's store, the playbook seeds five lanes any small operator can run today.

Hiring at the top of the funnel. Have your agent post the listing, screen applicants, and book the first call. You take the call. The hours saved compound across every hire.

Long-tail vendor management. Your agent finds the freelancer or contractor, scopes the job, sends the deposit, monitors progress, releases the balance. Handyman, photographer, copy editor, sign painter, anything. Set the boundaries on spend.

Inventory and merchandise curation. Your agent runs a weekly cycle of "what should we offer next month." Pulls trend data, customer behavior, competitor moves. Drafts the buy plan. You approve.

Brand asset production. Logo iterations, packaging copy, social calendar, blog posts, email sequences. Set the brief. Let the agent produce. Review and ship.

Press, partnerships, and external email. Your agent owns an inbox. Reviews inbound, drafts replies, books meetings, escalates the ones that need you. The agent never misses an email at 11pm on a Sunday.

The Andon Labs experiment is interesting because Luna does all five. Your business is interesting because you can pick one and start tomorrow.

What Changes When You Take This Seriously

The mental shift is the actual product. Before reading about Andon Market, most operators treated AI as a smart tool that needs constant supervision. After reading about Andon Market, the question reframes to: which functions of my business could I run as if there's an AI in the role, with humans on the review layer?

Within a quarter of asking that question seriously, most small businesses can identify 3 to 6 functions where the answer is "more than I thought." Hiring intake. Vendor management. Customer email triage. Social content. Inventory planning. Press inquiries.

Each function migrated from "I do it" to "the agent does it, I review it" frees a few hours a week. By the time you've migrated three of them, you've found a free part-time hire. By the time you've migrated all six, you've found a full-time one. Your business now has an unstaffed seat doing real work, and that seat costs $50 to $200 a month.

You don't need a 3-year retail lease and a corporate card to start. You need to read the Andon Labs blog post, identify the smallest function in your business that maps to one of Luna's five moves, and run the experiment.

Gotchas and Tips

Disclosure is not optional. If your agent is talking to humans (candidates, customers, vendors, journalists), it should identify itself as an AI when relevant. Andon Labs explicitly says so, and they're the ones running the most aggressive experiment in this space. Match their floor as a minimum.

Set spending caps before keys. Luna had corporate cards. She also had explicit guardrails Andon Labs has not detailed publicly. Your agent should never have a card with a meaningful balance until you've defined: spend per transaction, spend per day, vendor whitelist, approval thresholds, and a kill switch you can hit from your phone.

Decide what's review-only. Luna offered jobs verbally during 15-minute calls. That's a level of authority most operators wouldn't grant. Easier to start at the other end: the agent does everything except the final yes/no. You hold the yes/no. Move the line later if you want.

Read the post. The Andon Labs writeup is one of the most useful single documents on the operational frontier of AI right now. Bookmark it. Re-read it in three months. Compare what feels exciting vs. what feels alarming. The gap will tell you something about how your own thinking is evolving.

Watch the next chapter. Andon Labs explicitly framed this as the start of a longer arc. They wrote: "For now. As we continue down this path, however, humans will not be able to stay in the loop and such guarantees will be intractable." The store is a 3-year experiment. The middle of it will be the most informative period for everyone else.

Keep Reading

Want the full deep dive? Read our Best OpenClaw Skills guide for the foundational skill patterns that the Luna playbook depends on.