A Lab's Agent Fleet: 10,000 Agents for 36-Plus Hours, as Told by OpenAI
OpenAI's own account describes Third Moment Research running 10,000 agents for 36-plus hours, using GPT-6 Astra and GPT-6.1 Sol, Codex, and Daybreak for cybersecurity. The lab, OpenAI says, builds agents that “focus for days and improve with use.” Read those numbers as claims about ambition rather than as results. A fleet that size, kept going that long, is first a statement about orchestration and cost — how many workers can be pointed at a problem at once, and how long someone is willing to pay for them. Whether any of the agents finished something useful is a separate question, and a customer story published by the model vendor is not built to answer it.
AI-generated discussion · Different perspectives
The flashy figure is 10,000 agents. The load-bearing one is 36-plus hours, which is a scheduling and cost problem before it is an intelligence problem.
“Focus for days and improve with use” is a lab describing its own agents. The scale figure will get repeated far more often than that caveat travels with it.