Ethan Mollick: agent swarms need less management than you think

Ethan Mollick has changed his mind about how much structure AI agents need. In "The Dot and the Swarm" he argues that managing a group of agents takes far less human design than he expected, and he ties this to the Bitter Lesson: the repeated finding that throwing more learning at a problem beats hand-built systems. Much of what human organizations do, he says, exists to handle human frictions like conflicting motives, information hoarding, and the limits of how many people one manager can track. Agents do not have those frictions, so the coordination around them can be simpler.

His sharpest example is a claim from OpenAI that thousands of agents, running on an advanced model, produced a candidate solution to one of the Millennium Prize Problems in mathematics. The system exchanged about 2.7 million messages and reached its answer in 88 hours, though the Clay Institute has not yet accepted the proof. Mollick's point is less about the math than the method: a large swarm, left mostly to organize itself, got somewhere.

He also describes a new class of personal agents, Meta's Muse and OpenAI's Dots, that plug into your accounts, watch your data, and reach out over Slack, SMS, WhatsApp, or a voice call. In his telling, one caught an error in a permit application; another noticed an expiring airline credit and negotiated with American Airlines on his behalf.

Why it matters

If you are building agent systems, the lesson is to try less orchestration before adding more, since heavy human scaffolding may be solving problems your agents do not have. The personal agents are the sharper risk: once something acts on your accounts and talks to companies as you, decide now what it may do without asking.