A New Platform Lets AI Agents Run a Business

Andon Labs opened a platform this week called Pion, and the pitch is bolder than most. Rather than helping a person with a task, its AI agents are meant to run a business on their own. In the research preview announced on September 14, a company can hand day-to-day operations to a fleet of agents wired into the tools a real business uses, including corporate banking, email, phone lines, browsers and terminals.

The design puts a supervisor agent, named Andonos, at the top. A person gives it a high-level instruction, and it directs a set of sub-agents that do the actual work, the company explains. The structure is meant to keep a single overseer accountable for what the swarm below it decides to do.

Two years of vending machines

Pion did not arrive out of nowhere. Andon Labs spent close to two years moving from simulation to real money, and the story is oddly humble. It began with Vending-Bench, a test that had models manage an imaginary vending machine, then graduated to a physical one installed at Anthropic's office in early 2025. The agent lost money at first, tripped up by the mess of the real world, and only turned a profit late that year as the underlying models improved, as HyperAI recounts.

From there the experiments grew into Andon Market, a retail store in San Francisco, and Andon Cafe in Stockholm, both launched in April 2026. Neither has turned a profit yet, weighed down by fixed costs and the gap between an agent that can reason and one that can run a shop. The company is candid about that, which is more than most product launches manage.

The unresolved question is not whether this is clever. It is what happens when software with a bank account and a phone makes a bad call at three in the morning with no one watching. Handing agents real infrastructure is the whole point of Pion, and also the whole risk. A research preview is the right place to find out which one wins.