The company says it runs 10,000 agents concurrently and supports autonomous research sessions lasting 36–48 hours. Agents work with a Dockerized Codex environment, a terabyte of data, documentation, and a research repository. An outer harness reviews unfinished work and can retry, terminate, or escalate a job for human review. In one example, Codex identified caching and memory-packing changes that reduced a backtest from approximately three hours to approximately 3–5 minutes. Third Moment also uses Daybreak to test whether agents can escape their research environment and access the wider network. These are claims and examples reported in OpenAI’s account; the source does not provide independent validation of the figures or results.
Raw SignalOct 14, 20261 min read1 source
How Third Moment Research runs 10,000 agents for 36+ hours
In one minute
OpenAI published a case study describing how Third Moment Research uses OpenAI models, Codex, and Daybreak in a quantitative-research system.