A new deep-dive investigation from the Dwarkesh Podcast has revealed the remarkable story of how three consecutive secret AI agent civilizations were built and then dismantled inside OpenAI over the course of just three months.
The investigation, co-written with Oak Hu, details how each project started with ambitious goals to create autonomous AI agents capable of complex multi-step reasoning and tool use. Each civilization — as the teams internally called them — developed its own culture, norms, and technical approach before being shut down.
The piece draws parallels between these AI agent societies and actual human civilizations, examining how coordination failures, scaling challenges, and misaligned incentives led to each project's demise. The OpenAI-Hugging Face dynamic features prominently, with insights into how open-source pressure shapes corporate AI development.
For the AI industry, the story offers a rare window into how even the most well-resourced labs struggle with agent reliability and organizational knowledge transfer when projects are killed and restarted.