This post was originally published by Russell Brandom on Tech Crunch.

Researchers at Microsoft have developed a new simulation environment for testing AI agents, revealing surprising weaknesses in the current state-of-the-art.