How Third Moment Research runs 10,000 agents for 36+ hours

Third Moment Research is developing a self-improving AI system designed to handle company-specific workflows and continuously enhance its capabilities through use. By leveraging OpenAI models such as GPT-6 Astra, GPT-6 Sol, and GPT-6.1 Sol, along with Codex and Daybreak for cybersecurity, the AI lab is capable of running 10,000 agents simultaneously on investigations lasting between 36 and 48 hours. One of the company's early applications is quantitative finance, providing a controlled environment to study agent performance against clear metrics of success. In this setting, Codex was used to analyze simulation trajectories and identify a major backtesting bottleneck. After profiling the inputs, Codex deployed a caching optimization and a memory-packing technique that reduced backtest execution from approximately three hours to three to five minutes, thereby freeing up token budget for further experiments. To address agents that stop prematurely, Third Moment uses an outer harness powered by GPT-6 Sol or GPT-6.1 Sol to review incomplete tasks and decide whether to retry, terminate, or escalate. The agents operate inside a Dockerized environment equipped with a terabyte of data, a research repository, and tools to run feature engineering and model training. According to Adam, this progression in reasoning capability has allowed agents to explore unanticipated research directions, such as identifying the impact of non-displayed trades on an exchange. Furthermore, the company uses the Daybreak platform to automate penetration testing and ensure agents cannot break out of their research environments. Looking ahead, Third Moment plans to scale its operations to hundreds of thousands of collaborating subagents supported by trillions of tokens, aiming to automate complex research and optimization tasks.