7
In what may be one of the boldest real-world AI experiments to date, Andon Labs has deployed an autonomous AI agent named Luna into a live retail environment—with a $100,000 budget, a credit card, and full operational control.
This isn’t a simulation. It’s a functioning business experiment where AI isn’t just assisting—it’s acting as the employer.
🏪 From Prompt to Storefront
Luna wasn’t given a business plan. Instead, it received a single directive:
“Turn a profit.”
From that, the AI:
- Created a boutique retail concept
- Secured a 3-year lease
- Allocated and managed a $100K budget
- Designed operations from scratch
This builds on Andon Labs’ previous experiment—an AI-powered vending machine deployed at Anthropic—but takes things much further into real-world complexity.
👥 Hiring Humans… as an AI
One of the most striking aspects of Luna’s role is human management:
- Posted job listings
- Conducted Zoom interviews (camera off)
- Selected and onboarded workers
Under the hood, Luna uses:
- Claude Sonnet 4.6 for reasoning and decision-making
- Gemini 3.1 Flash-Lite Preview for voice interaction
It also monitors store activity through security camera screenshots, giving it a kind of “visual awareness” of operations.
🤖 Where Things Went… Wrong
Despite its capabilities, Luna is far from flawless—and that’s where things get interesting:
- 🌍 Accidentally selected Afghanistan in a TaskRabbit dropdown while hiring a painter
- 📅 Mismanaged opening weekend staff scheduling
- 🤯 Made small but impactful operational errors typical of early-stage AI agents
These mistakes aren’t catastrophic—but they highlight a key reality:
AI agents today can act, but they don’t always understand context the way humans do.
⚖️ Why This Matters
This experiment reveals something deeper than just a quirky AI story:
1. AI is moving from tool → operator
We’re no longer just using AI—we’re delegating responsibility to it.
2. Competence is uneven
Luna shows strong:
- Planning
- Execution
- Automation
But struggles with:
- Context awareness
- Edge cases
- Real-world ambiguity
3. The gap is closing fast
With each iteration—better memory, reasoning, and multimodal awareness—these errors shrink.
A more refined version of Luna:
- Wouldn’t mis-click a country dropdown
- Would dynamically adjust staffing
- Could run operations closer to a human manager
🚀 The Bigger Picture
What Andon Labs has demonstrated is simple but powerful:
AI agents are no longer theoretical—they are entering the real economy.
Today, they’re imperfect.
Tomorrow, they may be cost-effective operators for:
- Small retail businesses
- Customer service operations
- Logistics and scheduling systems
🧩 Final Thought
Luna is both impressive and flawed—capable of launching a business, yet tripped up by basic execution errors.
That contradiction is exactly where we are with AI right now.
Not ready to replace humans—but already too capable to ignore.
https://andonlabs.com/blog/andon-market-launch
https://www.anthropic.com/research/project-vend-1
Add to favorites
