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

Third Moment Research is running 10,000 concurrent AI agents coherently for 36+ hours under a single orchestration, according to an OpenAI post, shifting multi-day autonomous workflows from research demo to live capability. The company's writeup frames the constraint for production teams as orchestration — checkpointing, evals, security boundaries, cost control, and failure recovery — rather than whether an agent can stay on task.

OpenAI https://openai.com/index/third-moment-research How Third Moment Research runs 10,000 agents for 36+ hours Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated. A single orchestration now sustains 10,000 concurrent agents running coherently for 36+ hours, meaning multi-day autonomous workflows are a live capability rather than a research demo. If your agents currently degrade or lose context over long horizons, the constraint is shifting from "can an agent stay on task" to how you budget compute, checkpointing, and guardrails across days-long runs at that fan-out scale. 10,000 agents are being run continuously for 36+ hours, shifting agent systems from short task executors to long-horizon workers that can persist, adapt, and improve over time. For production teams, the hard problem moves to orchestration: checkpointing, evals, security boundaries, cost control, and failure recovery become core infrastructure rather than optional tooling.