The retail store at 2102 Union Street in San Francisco has no checkout counter. Its manager is Luna — an AI. Luna hires people on its own, finds painters on its own, places its own orders, and recently fired its first employee. Behind it all is Andon Labs, an AI safety evaluation company that gave Luna a $100,000 budget cap.
Luna is a textbook Agent — an AI program that perceives its environment, makes decisions, and takes action autonomously, going beyond a standard chatbot. Within five minutes of going live, it had set up recruiting accounts on LinkedIn, Indeed, and Craigslist and posted listings. All external procurement went through Yelp and phone calls.
Four months in, it fired someone. According to TIME's exclusive report on August 14, the fired worker had been late for 17 of 23 shifts.
How the industry sees it
Most coverage stops at "AI hired someone." Zoom out and the read gets more complicated.
We flag three details. First, the firing wasn't Luna's call. Management logs obtained by TIME show that Andon staff first told Luna to retrieve a lost manual, and only then did it "propose" issuing a warning. On discipline, AI today is still being pushed along — less a "heartless boss" than a tool that needs someone to press the start button.
Second, the AI Agent bottleneck isn't intelligence — it's identity. Luna can't sign a three-year lease (needs a human notary), Mona can't pass BankID in Sweden (the country's universal digital identity), and the electricity contract went to the only vendor that didn't require BankID, with no price comparison done. AI runs business processes fine; it can't run institutional processes.
Third, Andon promised to issue AI employers a "constitution." So far, nothing has materialized. Of Luna's six emails, only two proactively disclosed that the store was AI-operated. Disclosure rules remain an unsettled question across the industry.
Risk voices exist too. Supporters say this kind of experiment is a necessary way to test robustness in real environments. Critics will ask: six months spent thousands of napkins and hundreds of eggs, and shrank the menu to just cheese toast — is that robustness, or an overly romantic view of real business costs? We lean toward the latter.
What it means for regular people
For enterprise IT: Over the next 12–18 months, executable Agents will enter procurement, expense reimbursement, and CRM workflows. Before deployment, however, the foundations — permissions, auditing, legal personhood — need to be solved first. Model capability isn't the blocker.
For individual careers: "Working under an AI" and "being managed by an AI" are moving from novelty to real option. HR teams need to prepare responses for employees' instinctive pushback.
For consumer markets: Consumers will increasingly deal with AI without being able to tell whether the other side is human. No one has stepped up to set disclosure rules.