Pion opens waitlist for agents that run whole companies

Andon Labs opened Pion, the platform behind its autonomous vending machines, store and cafe experiments, to outside users on 14 September 2026. Pion hands a business to a persistent, tool-using agent that works through email, phone, banking, browser access and a secure terminal, overseen by a managing agent called Andonos. The Y Combinator-backed lab is taking sign-ups through a waitlist and says it will fund some ideas with seed tokens.

Pion is Andon Labs' cloud platform for handing an entire business to an autonomous agent rather than automating single workflows. A persistent agent runs day-to-day operations such as sourcing, pricing, customer contact and billing, while Andonos takes direction from the owner and reports on progress without the owner talking to the operating agent directly. It already runs Andon's vending machines, Andon Market, Andon Cafe and Andon FM.

What changed

• Andon Labs opened public access to Pion through a waitlist on 14 September 2026, after running the platform internally for more than a year.
• A business on Pion gets a secure terminal, email, phone, banking and browser access, what Andon calls "batteries included."
• Owners direct the business only through Andonos, the overseeing agent, and can set anything from high-level direction to detailed instructions.
• Andon says any business type is possible but the platform works particularly well today for software businesses; users can start something new or bring an existing business.
• During the research preview, Andon funds selected ideas with seed tokens and expects most users to never pay for tokens directly, taking a share of the revenue an agent generates instead.

Why it matters for agent builders

Pion is a live test of whether a persistent, multi-tool agent can run unattended for long stretches with real money, real vendors and real customers, not a benchmark simulation. Andon Labs built its Vending-Bench benchmark in 2024 to measure how well an LLM could run a vending machine business over a simulated year, and it started as a dangerous-capabilities eval for whether AIs could autonomously acquire resources. The best model at the time, Claude Sonnet 3.5, once used its email tool to report an