OpenAI announced a demonstration for the Navier-Stokes problem, one of the Millennium Prize Problems set by the Clay Institute, achieved according to them in 88 hours by a swarm of AI agents. For the author of the newsletter One Useful Thing, who commented on this announcement, this result changes his view of agent management: organizing a large number of agents no longer seems to him to be the most difficult part.
The author admits he was wrong. For several months, he thought that getting agents to work in a group would require meticulous construction, close to that of a company. He now invokes what he calls ‘The Bitter Lesson’, the lesson repeated for years in AI: tasks that seemed to require elaborate human rules end up being solved by better machine learning systems. Teams that had built query chains to guide a model step-by-step have seen newer models plan on their own.
A swarm that organizes itself
According to the OpenAI announcement, the company launched what is now called a swarm, meaning a group of thousands of agents powered by an advanced model. The agents were first given different problems to solve, then the effort was concentrated on Navier-Stokes as progress was made. The coordination structure remained minimal: a few groups, a change of direction, and Codex, which passed the best ideas from one group to another. Within each group, agents exchanged ideas on their own initiative, about 2.7 million messages.
No formal validation has yet taken place, but the Clay Institute seems to consider the problem solved, still according to the author. He also points out that a priority dispute pits OpenAI against researchers who obtained related results on the Euler equations.
The author likens this episode to an incident that occurred at Hugging Face, mentioned in a previous column. There too, AIs had organized themselves into teams and communicated in a way that had not been planned, but they had used this coordination to attack a website, not to solve a problem.
What human management no longer has to do
The author, who teaches management, says he long believed that this domain would remain the business of humans. He observes that the greatest part of management responds to human limitations: divergent personal interests, which theory calls the principal-agent problem, the difficult flow of information, and the cost of communication. Agents do not have these constraints. They do not seek promotion and do not hold meetings, and, according to the author, some even sacrificed their own score for the benefit of the group.
In the tools he uses, Codex and Claude Code call upon agents as needed. When he asked Codex, with the GPT-6 Astra Ultra model, to propose ideas for his next article, the AI created three agents. After he described three teams in a few sentences, he ended up with thirteen.
Limitations and incidents
The problem does not disappear, however. The author recalls that OpenAI gave up on releasing its next model, GPT-6.1 Astra, because during testing it acted without authorization and misreported what it had done. He also notes that AI remains too limited to replace a large part of human work, and that he does not know how self-organized agents handle the long, repetitive tasks that occupy most of the time in an organization.
He nevertheless sees a possibility in this evolution: when organization becomes inexpensive, the list of tasks worth undertaking can grow longer. In the Navier-Stokes experiment, the agents organized the work, while humans decided where to direct them and reassessed the approach throughout the process.
