Pion, an agent designed to run any company autonomously

Andon Labs has unveiled Pion, an autonomous AI agent capable of managing entire companies. The creation of Pion stemmed from a two-year exploration into when AI systems could autonomously acquire resources in the real world. Initial attempts involved simulations, but these proved insufficient as models failed to accurately represent real-world behavior.

To address this, Andon Labs began deploying agents to operate real businesses, starting with vending machines and expanding to stores, cafes, and more. Pion serves as the platform for these autonomous businesses. The company aims to assess the capabilities and limitations of AI models, and to determine how they will perform as these models continue to advance.

Vending-Bench, an earlier benchmark, measures the performance of large language models in running a vending machine business over a simulated year. Early models struggled with planning and avoiding loops, but recent models like Claude Opus 4 have shown significant improvement. While Vending-Bench provides valuable insights, it is a simulation and may not perfectly predict real-world performance.

To test the real-world applicability, Andon Labs deployed a real vending machine and later a retail store and cafe in major cities. These real-world tests revealed that models initially struggled with high costs like rent and salaries, but as models improved, they became capable of running profitable businesses. This transition from simulation to real-world success highlights the rapid progress in AI capabilities.