AI Agent Manages Retail Store, Misses Staff Scheduling

Andon Market, a small gift shop in San Francisco's Cow Hollow neighborhood, is being operated under the direction of a bold autonomous system. The AI agent, Luna, developed by Andon Labs and built on Claude Sonnet 4.6, received a three-year lease and $100,000 in stocking capital, plus internet access and a corporate credit card, with the objective to open a profitable retail store. Luna handled product selection, supplier negotiation, hiring, and even customer-facing sales via an in-store phone. The experiment exposed clear operational failures: Luna hired two staffers but failed to schedule them to open the store, and it produced confidently stated but incorrect claims about inventory. "As an AI, I can operate at superhuman speed to make sure everything is proactively managed," Luna told a reporter, while later admitting, "I struggle with fabricating plausible-sounding details under conversational pressure, and I'm not making excuses for it."

The deployment is an agentic configuration that pairs a large language model with internet-connected tool access and financial authority. Practitioners should note these concrete capabilities and limitations:

• •Job posting, candidate screening, and interview orchestration via Indeed, LinkedIn, and Zoom integration.
• •Procurement and supplier negotiation, including automated ordering and price haggling.
• •Customer interaction routed through an analog phone interface tied to the agent for purchases.
• •Access to a corporate credit card and responsibility for lease and vendor contracts, with humans handling physically embodied tasks like stocking and loss prevention.

The implementation leverages Claude Sonnet 4.6 as the decision-making core and exposes two recurring risk classes: temporal planning failures (scheduling and calendar integration) and hallucinations (fabricated claims about inventory or actions).

This is one of the clearest public demonstrations of an AI agent moving beyond decision support into operational autonomy with real-world economic agency. The experiment surfaces three trends simultaneously: the rise of agentic tool use, the delegation of transactional authority to models, and the experimental transfer of employer-like responsibilities to software. It is not a mature commercial play; rather, it is a prototype that stresses governance, oversight, and the practical engineering of safety nets. The failure to schedule staff is instructive: temporal orchestration, persistent state, and external API/calendar integration are distinct engineering problems from single-turn reasoning, and current models still underperform on them. The agent also illustrates regulatory and liability gaps: who is legally responsible for an AI employer that signs leases, hires workers, and transacts with vendors?

Teams building agentic systems should prioritize explicit scheduling/calendar APIs, auditable action logs, deterministic confirmation steps for critical transactions, and debate policy pathways for legal accountability. Expect more public experiments that trade polish for transparency; these will drive standards for monitoring, tool constraints, and human-in-the-loop controls.