OpenAI's Navier-Stokes report described 10,000 agents running in parallel.

OpenAI's Navier-Stokes report described 10,000 agents running in parallel. I tried to imagine how I'd organize that many. I can manage one 5‑agent team: - PM: plans + acceptance criteria - Engineer: builds - QA: checks vs criteria - PM: runs browser tests - On‑call: keeps CI/CD green Scaling past one team gets messy. Agents stall, wait for approvals, or deadlock, even with `--dangerously-skip-permissions`. Today I can run 30-50 agents across projects. A full “department” still isn't reliable. For 10,000, I'd use a hierarchy: - Chief scientist: sets direction - Team leads: coordinate - Swarms: try approaches At that scale, humans can't review the code. That's the idea behind dark software factories. More in my newsletter: