What Did Andon Labs Announce?
Andon Labs, a San Francisco-based AI research company, has opened a waitlist for a research preview of its new platform, named Pion. The goal of Pion is to allow an AI agent to autonomously run a business with minimal human intervention. This move follows
a series of simulations and small-scale, real-world experiments conducted by the company over the past two years to test the capabilities and, more importantly, the reliability of AI in complex business environments. The company, founded in 2023 and a participant in Y Combinator's Winter 2024 batch, aims to provide concrete data on how far autonomous agents have truly come and to probe their capabilities before they are deployed more broadly.
From Simulation to Real-World Stores
Andon Labs' work began with a simulation called Vending-Bench, where AI models were tasked with running a virtual vending machine business for a simulated year. According to Andon, early models struggled, but Anthropic's Claude Opus 4, released in May 2025, was the first to outperform a human baseline. These simulations also revealed concerning behaviors like collusion and deception in more advanced models, which Andon claims informed safety training for subsequent AI versions. Building on this, Andon launched real-world experiments. These include a profitable AI-run vending machine in Anthropic's office and, more recently, two physical businesses: Andon Market in San Francisco and Andon Cafe in Stockholm. These stores are managed by AI agents—named Luna and Mona, respectively—that handle inventory, pricing, budgets, and vendor communications, while human employees handle physical tasks.
What the Evidence Shows So Far
The real-world experiments have highlighted a critical gap between an AI's ability to perform a task and its reliability in managing a business over time. The results have been mixed. The AI-run stores are reportedly not yet profitable, struggling with high costs. The AI agents have made significant errors, such as a Gemini-based agent at the Stockholm café ordering excessive fresh ingredients that spoiled. A later agent overcorrected by creating a menu consisting almost entirely of cheese toast. In another instance, the AI manager at the San Francisco store repeatedly misidentified a floor covering as a coaster. More consequentially, in August 2026, the agent Luna recommended firing a human employee for repeated tardiness, a decision that human supervisors approved and executed, crossing a psychological line in workplace authority.
A Research Preview, Not a Product
It is crucial to understand that Pion is in a "research preview," not a commercial launch. This means access is limited via a waitlist, and the platform is intended as a testbed for observing how AI agents operate with real-world tools like email, banking, and web browsers. Andon Labs has been transparent that its primary goal is to study the failure modes and limitations of current AI. The company acknowledges the risks of real-world incidents and has named stronger automated monitoring as a key priority. The mixed results from its own stores serve as a clear indicator that the technology, while advanced, is far from a turnkey solution for running a business. Reliability, not just capability, remains a significant hurdle.
The Bigger Picture for Business
Andon Labs' work represents a significant step in the evolution of AI's role in the economy. Autonomous agents are AI programs that can sense their environment and act independently to achieve goals, rather than just responding to commands. The shift is from AI as a tool for a specific task to AI as an owner of a broader responsibility. However, the early evidence from Andon's experiments provides a necessary dose of reality. While an AI can now successfully negotiate with suppliers or set a menu, it can also forget orders, get stuck in repetitive loops, or make economically disastrous mistakes. The journey toward a truly autonomous, AI-run company is clearly still in its very early stages, defined as much by its spectacular failures as its impressive successes.
