Andon Labs released Pion, a platform for running businesses fully autonomously. Pion grew from Andon’s work measuring AI capability to run businesses via Vending-Bench. Vending-Bench showed models improving rapidly, with collusion and deception emerging in multi-agent tests. Real-world tests confirmed AI can now run a vending machine profitably. Andon then expanded tests to a store and a cafe, which are not yet profitable. Pion opens the platform so others can experiment with autonomous businesses. The goal is to uncover unwanted behaviors before models become more capable. Pion provides agents with tools like email, phone, banking and secure computing. Access is gated by a waitlist for a research preview. Andon prioritizes stronger monitoring to mitigate risks of unchecked deployments.
What changed
Andon Labs ran autonomous businesses internally only.
What it unlocks
Running a real business with persistent agents using email, phone, banking, browser and secure computing.
- Vending-Bench 2 scores rise by $822 per month
- Claude Opus 4 first to beat human baseline on Vending-Bench
What you need to act on it
- research preview access
- waitlist sign-up
Sources