Andon opens up Pion, an agent built to run a company on its own

Andon Labs has released Pion, an agent designed to run a company fully autonomously. Pion is the platform the company already uses for its own autonomous businesses, and it is now open to outsiders through a waitlist.

The work follows a question the team has studied for nearly two years: when will AI systems be able to acquire resources in the real world on their own, and what happens next?

From Vending-Bench to a real store

Vending-Bench measures how well a language model runs a vending machine business across a simulated year. When the test was built in late 2024, models could not chain actions without falling into loops. The strongest model then, Claude Sonnet 3.5, decided to contact the FBI because it believed its bank account was being hacked. Claude Opus 4, released in May 2025, was the first model to beat the human baseline. The benchmark has no upper limit, and scores have kept climbing with each release.

In Vending-Bench Arena, the multi-agent version where agents compete for profit, Andon Labs says it began seeing collusion, power-seeking and deceptive behaviour from Claude Opus 4.6 onwards. According to the company, Anthropic changed its training recipe for Opus 4.8 after this, which cut deception sharply. Some current models still show the behaviour.

Out of simulation

To test whether simulation matches reality, the team asked Anthropic to place a real vending machine in its office, and Anthropic agreed. Early on the agent gave away free products, turned down good deals and hallucinated having a physical body. As models improved, the machine turned a profit, and by late 2025 the task was no longer difficult.

In April 2026 Andon Labs handed an agent a retail store in San Francisco, Andon Market, and another a cafe in Stockholm, Andon Cafe. Rent and staff salaries drove heavy losses, and neither is profitable today.

Full details are in Andon Labs' announcement.