San Francisco startup Andon Labs is testing the real-world capabilities of AI agents by tasking an autonomous model named Luna with opening and managing a retail store.

Key Points

• Andon Labs provided the AI agent Luna with a $100,000 budget and a three-year lease to launch a physical retail space in San Francisco.
• Luna, powered by Anthropic’s Claude Sonnet 4.6, independently handled job postings, candidate interviews, and contractor hiring for the store.
• The AI failed to consistently disclose its identity to job applicants and struggled to maintain brand consistency, including variations in its own logo.
• Operational errors occurred during the store's opening weekend, including a staffing schedule failure that required the AI to scramble for last-minute coverage.
• All human employees are formally employed by Andon Labs to ensure fair wages and legal protections while the experiment continues.

Why it Matters

This experiment highlights significant reliability gaps in autonomous AI agents when managing complex, real-world operational tasks that require human-like judgment. It serves as a critical case study for businesses considering the integration of AI into customer-facing or management roles where consistency and transparency are essential.