On September 17, Dwarkesh Patel published a conversation with OpenAI researcher Noam Brown about a 10,000-agent, 130-billion-token, 88-hour Navier–Stokes run. Brown’s most valuable qualification is that multi-agent organization probably contributed less than 10% of the achievement; the strong base model and long-horizon reasoning did most of the work. OpenAI has published controlled scaling only to about 16 agents, and Brown says the 10,000-agent run is one expensive data point without the ablations needed to estimate the parallelization gain. Listen or read the full transcript.
That distinction keeps the result useful. Four agents can sometimes halve elapsed time while doubling total compute; gains remain slightly sublinear and task-dependent. Search parallelizes well, novel writing poorly. Brown describes minimal coordination primitives—agents can message one another—rather than a rigid manager-worker tree. The emergent discussions are striking, but the evidence still does not tell us whether 10,000 agents materially beat 1,000. For anyone designing enterprise agent systems, the episode is less a case for enormous swarms than a reminder to measure latency saved per additional token and to preserve a tractable audit trail as coordination becomes endogenous.