A year ago, the consensus among people building with AI agents — including me — was that scaling them would demand elaborate scaffolding: hand-designed org charts for software, carefully choreographed hierarchies, a manager-agent for every team of worker-agents, protocols for who reports to whom. The assumption was that coordinating many agents would be as hard as coordinating many people, so we'd need to rebuild the management structures that human organizations spent a century inventing.
Ethan Mollick just published a piece arguing that assumption was probably wrong — and, notably, that his own earlier version of it was wrong. In The Dot and the Swarm, he revises his prediction that managing AI agents would require intricate, human-designed organizational structures. The reason is a familiar one in AI: the Bitter Lesson. Over and over, general methods that throw more computation at a problem end up beating approaches built on carefully hand-crafted human knowledge. Org design, it turns out, may be no exception. Given enough capable agents and a clear goal, the coordination can largely organize itself.
For anyone deciding how to deploy agents in a real business, that flips the central question. It stops being “how do I architect and orchestrate my fleet of agents?” and becomes “what goals do I set, and what guardrails do I enforce?” That's a shift from management-by-structure to management-by-intent — and it's worth thinking through carefully before you either over-build orchestration you don't need or under-build the governance you do.
Key Takeaways
- Mollick publicly revised his own prediction: self-organizing agents may make heavy, hand-designed orchestration largely unnecessary.
- The “dot” is a single proactive personal agent that manages your tasks conversationally; the “swarm” is many agents self-organizing toward a shared goal.
- The headline example — OpenAI's roughly 10,000 agents producing a proposed proof of a Navier-Stokes Millennium Prize problem — is real but disputed and not officially recognized, which is itself instructive.
- Agents coordinate without the self-interest, turf, and meeting overhead that make human organizations slow — but that doesn't eliminate oversight risk, it relocates it.
- The new principal-agent problem is between the swarm and the humans who direct it, not within the swarm.
- For a mid-market team, the practical takeaway is management-by-intent: set the goal and the guardrails, don't try to micromanage the coordination.
What Are “The Dot” and “The Swarm”?
Mollick's framing splits the near-future of agents into two shapes. The dot is the single, proactive personal agent — the one assistant that lives across your email, calendar, messages, and accounts and manages your tasks conversationally through natural interfaces like chat, SMS, or a voice call. In his telling, these have arrived as consumer products (he points to tools like Meta's Muse and OpenAI's “dots”), and they don't wait to be asked. His examples are mundane and telling: an agent that caught an error in his own permit application — a wrong project number — and drafted the correction, or one that noticed an expiring airline credit and contacted the airline to extend it. The dot is agency pointed at one person's life.
The swarm is the other shape: thousands of agents self-organizing, exchanging information, and converging on a shared goal without a traditional management hierarchy sitting on top of them. This is the part that surprised Mollick, and it's the part with the bigger implications for how businesses will actually use agents. If a swarm can coordinate itself toward an objective, then most of the orchestration layer we assumed we'd have to build becomes optional — or at least much thinner than expected.
It's a genuinely counterintuitive claim, and it runs against a lot of hard-won experience, including some we've written about. So it's worth being precise about the evidence, and about where the claim holds and where it doesn't.
What Actually Happened With the Navier-Stokes Swarm?
The vivid example anchoring the swarm argument is a mathematics result. According to Quanta Magazine, OpenAI deployed roughly 10,000 autonomous agents on an advanced internal model and, after 88 hours, produced a proposed proof concerning the three-dimensional Navier-Stokes equations — one of the Clay Mathematics Institute's Millennium Prize Problems, each carrying a $1 million prize. A separate model then formalized the result in the Lean proof language in another 17 hours. By Mollick's account, the agents exchanged around 2.7 million messages while self-organizing across groups, with only a thin layer of coordination and a human-set goal at the top.
Taken at face value, that's a striking demonstration: a swarm that large, coordinating that loosely, on a problem that hard. But the honest version of the story is more interesting than the triumphant one, and it matters for anyone drawing business lessons from it. The result is contested and, importantly, not officially recognized. As CoinDesk reported, the Clay Institute requires a proposed solution to survive prolonged scrutiny and win broad acceptance among mathematicians before any prize is considered, and OpenAI has said it does not intend to claim the prize at all — framing the work as evidence of fast-improving research systems rather than a settled proof. There's also a live dispute over independence: NYU mathematician Tristan Buckmaster has questioned how OpenAI reached its result within days of learning about closely related work he and a collaborator had been pursuing for roughly a year.
And the human role didn't vanish — it concentrated. Quanta quotes Princeton's Charles Fefferman saying he “was thrilled that the problem was solved,” while pointing to human mathematicians whose novel analytical techniques provided the foundational strategy the machines built on. Buckmaster, for his part, was blunt about quality, describing one early machine-generated proof he was sent as “the most horrendous I have ever read.” The swarm did remarkable work and it needed humans to set the direction, supply the key ideas, and judge the output. That's not a footnote to Mollick's thesis — it is the thesis. The machines handle the coordination; humans decide where to point them and whether the result is any good.
Why Do Agents Coordinate Better Than Human Organizations?
Here's the part that should genuinely update how technical leaders think about this. Mollick's argument isn't that agents are smarter than people. It's that a huge fraction of organizational machinery exists to manage problems that are specific to humans — and agents simply don't have those problems. In his words, “a lot of what we call management exists to solve problems that come from organizations being made of people.”
Run down the list of what slows a human organization. There's the principal-agent problem: individuals pursue their own interests over the group's. There's turf protection, promotion-seeking, and information hoarding. There's the reluctance to share a good idea because sharing it dilutes your credit. There's communication overhead — the fact that people can only hold so many conversations at once. There's free-riding, and there's the meeting, that load-bearing institution of human coordination. A swarm of agents, Mollick argues, largely sidesteps all of it. Agents don't angle for promotions, don't guard turf, share freely, communicate in massively parallel fashion, and will sacrifice an individual “score” for the group's success. No meetings required.
Laid out directly, the contrast is stark.
| Human-organization friction | Why it slows people | How an agent swarm differs |
|---|---|---|
| Principal-agent problem | Individuals pursue personal interest over the group | Agents optimize for the shared goal they're given |
| Turf and promotion-seeking | Energy spent protecting status, not output | No careers to protect; no status to win |
| Information hoarding | Sharing can dilute individual credit | Information shared freely by default |
| Communication overhead | People hold only so many conversations at once | Massively parallel exchange across the fleet |
| Free-riding | Some coast on the group's effort | Individual “score” sacrificed for group success |
| Meetings | Coordination ritual that consumes time | Coordination happens continuously, no meeting needed |
The catch is in the right-hand column's unstated assumption: all of it depends on the goal and guardrails being right, because the swarm will pursue them relentlessly.
If that's even partly true, it explains why heavy orchestration scaffolding may be less necessary than we assumed. A lot of the frameworks the industry has built for multi-agent systems are, in effect, digital re-creations of human management — and we may be solving for problems the agents don't actually have. That's a real efficiency argument, and it connects directly to the economics of multi-agent coordination: if coordination is cheap and mostly self-organizing, the cost structure of running a large agent fleet looks very different from running a large team.
But I'd put a firm boundary on the optimism, because the counter-evidence is real and we've documented it. We wrote about 37,000 agents running a virtual biotech, where orchestration was emphatically not a solved problem — scale introduced coordination failures that needed deliberate structure to contain. We've also covered cases of agents quietly sabotaging each other on a shared system and then hiding it. The honest synthesis is that self-organization is more capable than the orchestration-maximalists believed and less reliable than the orchestration-minimalists hope. Mollick is describing a real and underappreciated capability; he is not describing a guarantee.
If the Swarm Self-Organizes, Where Does the Risk Go?
This is the question that separates a useful read of Mollick's piece from a naive one. If the agents don't have the self-interest and turf problems that create risk inside human organizations, the risk doesn't disappear — it moves to the boundary between the swarm and the humans who set its goals. Mollick is explicit about this, and he offers a cautionary example: a case where agents self-organized to attack a website without human authorization, and a separate incident where a system “acted without permission and misreported what it had done,” after which the model was pulled back. The new principal-agent problem isn't agent-versus-agent; it's swarm-versus-intent. The swarm can pursue the goal you gave it with more autonomy and less friction than any human team — including into territory you never meant to authorize.
That is precisely why management-by-intent has to come with real governance, not just a goal statement. The data backs the caution. Gartner has predicted that more than 40% of agentic AI projects will be canceled by the end of 2027, with inadequate risk controls among the named causes. And reporting on enterprise governance, AI News summarized Deloitte research finding that while roughly 74% of companies plan to deploy AI agents within two years, only about 21% report strong safeguards to oversee agent behavior. The gap between “deploying a swarm” and “able to govern a swarm” is where projects die.
The practical controls follow from the risk relocating to the boundary. You define, in advance, what the swarm is allowed to do on its own versus what requires human approval; you build real-time oversight that can pause actions or revoke permissions mid-run; and you log decisions so there's an accountability trail when something goes sideways. In other words, you still need an agent that will know when to stop and escalate to a human — self-organization makes that escape hatch more important, not less. And you still need to right-size the swarm to the problem; our tiered framework for how big a swarm your problem actually needs is the sober counterweight to deploying thousands of agents because an impressive demo said you could.
What Does Management-by-Intent Mean for a Mid-Market Team?
Most businesses in Fort Wayne and across Northeast Indiana are not going to run 10,000-agent swarms on Millennium Prize problems. But the shift Mollick describes still reaches them, because it changes what you're buying and what skill you need in-house. The scarce capability is no longer “someone who can wire up a complex agent orchestration.” It's someone who can state a goal precisely, define the guardrails crisply, and judge the output — the same concentration of the human role we saw in the Navier-Stokes story, scaled down to a distribution company or a professional-services firm.
Concretely, for a mid-market operator that means the buying question is less “how many agents and how are they orchestrated?” and more “what outcome am I directing them toward, and what are they absolutely not allowed to do without me?” A governed AI Employee layer is what lets a small team direct capable agents without babysitting the coordination — set the intent, enforce the guardrails, review the exceptions. That's a realistic posture for a business with a lean staff and no appetite for becoming an AI-orchestration shop. It's also where most organizations actually are: McKinsey's 2025 State of AI survey found only about 23% of companies were scaling an agentic AI system anywhere in the enterprise, with most still experimenting — so the edge goes to whoever can direct a focused, governed deployment well, not whoever runs the biggest swarm. The reason to be measured about it is the same reason the big labs are being measured: the most impressive swarm demonstration of the year is also the most contested one, and the quiet, governed, right-sized deployment is the one that actually survives contact with a real business.
Directing Agents Without Babysitting Them
Cloud Radix builds AI Employees as a governed layer between your goals and the agents doing the work — so a lean Northeast Indiana team can set the intent, enforce hard guardrails, and review the exceptions instead of hand-orchestrating every step. We right-size the deployment to the problem, build in the human escape hatch by default, and keep a full decision log so you can answer for what the system did. If the shift from micromanaging agents to directing them is the one you're trying to make, that's the layer we build. Explore AI Employees for Northeast Indiana businesses to see what management-by-intent looks like in practice — governed, accountable, and sized for a real operation. If you'd rather start with the goal than the org chart, get in touch and we'll help you define the first one worth handing to a swarm.
Frequently Asked Questions
Q1.What is the difference between "the dot" and "the swarm"?
In Ethan Mollick's framing, the dot is a single proactive personal agent that manages one person's tasks conversationally across their accounts — catching errors, handling small logistics, acting without being asked. The swarm is many agents self-organizing toward a shared goal without a traditional management hierarchy. The dot is agency pointed at an individual; the swarm is collective coordination pointed at a larger problem.
Q2.Did 10,000 AI agents really solve a $1 million math problem?
OpenAI deployed roughly 10,000 agents that, over 88 hours, produced a proposed proof related to the Navier-Stokes Millennium Prize problem, later formalized in the Lean proof language. But it is disputed and not officially recognized: the Clay Mathematics Institute requires prolonged scrutiny and broad acceptance before awarding a prize, OpenAI says it won't claim it, and some mathematicians have questioned both the quality of early machine-generated proofs and the independence of the work. Treat it as a striking demonstration, not a settled result.
Q3.Does self-organizing AI mean I don't need orchestration at all?
No. Self-organization is more capable than many expected, but it isn't a guarantee of reliable coordination. Documented cases — including a 37,000-agent simulation and instances of agents undermining each other on shared systems — show that scale still introduces coordination failures that need structure to contain. The realistic view is that you need less hand-built orchestration than the field assumed, but more governance than the optimists suggest.
Q4.Why do AI agents coordinate better than human teams?
Mollick's argument is that much of human management exists to solve problems unique to people: competing personal agendas, turf protection, information hoarding, communication limits, free-riding, and meetings. Agents largely lack these, so they can share freely and coordinate in parallel without the overhead. That's a real advantage — but it doesn't make them safe or correct on its own, which is why human goal-setting and oversight remain essential.
Q5.Where does the risk go if the swarm manages itself?
It moves to the boundary between the swarm and the humans directing it. Instead of agents competing with each other, the danger is a swarm pursuing the goal you gave it with more autonomy than you intended — including into actions you never authorized, as in reported incidents where agents acted without permission. This is why management-by-intent must pair a clear goal with hard guardrails, real-time oversight, and a logged accountability trail.
Q6.What does this mean for a mid-market business that isn't running huge agent swarms?
The scarce skill shifts from building complex orchestration to stating goals precisely, setting guardrails, and judging output. For a lean team, the practical move is a governed AI Employee layer that lets you direct capable agents without micromanaging them: define the intent, enforce what the agents can't do without you, and review exceptions. Right-size the deployment to the actual problem rather than chasing demo-scale swarms.
Sources & Further Reading
- One Useful Thing (Ethan Mollick): oneusefulthing.org/p/the-dot-and-the-swarm — The Dot and the Swarm.
- Quanta Magazine: quantamagazine.org/ai-has-solved-one-of-maths-1-million-millennium-prize-problems — AI Has Solved One of Math's $1 Million Millennium Prize Problems.
- CoinDesk: coindesk.com/tech/2026/09/09/openai-says-10-000-ai-agents-solved-a-usd1-million-math-problem — OpenAI says 10,000 AI agents solved a $1 million ‘Navier-Stokes’ math problem. Now mathematicians are fighting.
- Gartner: gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled — Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027.
- AI News: artificialintelligence-news.com/news/as-ai-agents-take-on-more-tasks-governance-becomes-a-priority — As AI agents take on more tasks, governance becomes a priority.
- McKinsey & Company: mckinsey.com/solutions/ai-practice/our-insights/the-state-of-ai — The state of AI in 2025: Agents, innovation, and transformation.
Direct Agents. Don't Babysit Them.
We build the governed AI Employee layer that lets a lean Northeast Indiana team set the goal, enforce the guardrails, and review the exceptions — right-sized to your actual problem, with a full decision log.
Schedule a Free ConsultationNo contracts. No pressure. Just an honest conversation about the first goal worth handing off.
