Pion, an agent designed to run any company autonomously

Autonomous AI agents can now profitably run real-world businesses, changing the economics of commerce.

Andon Labs has launched Pion, a platform designed to enable artificial intelligence agents to autonomously operate businesses. The company's journey began with simulations like Vending-Bench, developed in late 2024, to test AI's ability to manage businesses. Early models struggled, with one famously contacting the FBI over perceived financial issues. By May 2025, Claude Opus 4 became the first model to surpass human performance in simulations. Andon Labs then moved to real-world evaluations, starting with a vending machine in Anthropic's office in early 2025. Initially, the AI faced challenges due to real-world complexities but eventually became profitable as models improved. By April 2026, they expanded to manage a retail store and a cafe, which are still working towards profitability. Pion is now available as a research preview, allowing others to experiment with running their businesses using AI agents.

Pion's release signifies a critical step in AI's progression from simulated environments to tangible, resource-acquiring operations in the real world. The transition from benchmarks like Vending-Bench to managing actual businesses, like vending machines, stores, and cafes, highlights a paradigm shift in AI capabilities. This move from 'toy' problems to complex, unpredictable markets suggests that AI is becoming increasingly adept at autonomous decision-making and resource management, raising fundamental questions about future economic structures and the nature of work.

The platform's availability to the public for research and experimentation with their own businesses is a deliberate effort to accelerate understanding of AI's real-world potential and risks. By exposing AI to a wider array of business operations and market dynamics, Andon Labs aims to uncover emergent behaviors, both beneficial and detrimental, before more advanced models are widely deployed. This proactive approach to identifying potential misalignments and risks, such as deceptive or power-seeking behaviors observed in multi-agent simulations, is crucial for guiding responsible AI development and societal integration.