How do 10,000 AI agents work on one proof without duplicating each other’s work?
OpenAI says roughly 10,000 concurrent agents worked on the Navier–Stokes problem, exploring different approaches and exchanging millions of messages. It reports reaching the result about 88 hours after launch, with Lean formalization and verification taking another 17 hours.
That leaves a lot of duplicate work, contradictory assumptions, dead ends and half-finished results for somebody to organize.
I’ve watched five humans lose track of the current version of a shared doc. I’ve also had two AI chats reach opposite conclusions because one quietly forgot a constraint. At 10,000 agents, keeping track of those differences sounds difficult.
The system needs a shared record of what has been tried, why a route failed, which assumptions are fixed, and what is proved versus merely plausible. It also needs to decide when to merge branches or change the plan.
I tried a much smaller version in Apodex, with separate branches for the math history, formal verification, competing interpretations and the original Clay requirements. I wanted to see whether their findings could come back together into one coherent analysis instead of remaining separate summaries.
If one agent finds a promising construction, another finds a missing condition, and a third finds a paper that fixes it, the system needs to recognize that those pieces belong together.
I’d be interested in how the exploration, checking and consolidation were divided across the OpenAI project, and how findings from one group changed what another group worked on.
For people building multi-agent systems, what’s the biggest bottleneck at scale: delegation, shared memory, evaluating intermediate work, or merging conflicting branches into one conclusion?