Agents Are Drawing Their Own Org Charts. Humans Still Set the Goal

In The Dot and the Swarm, Ethan Mollick revises his own assumption: he expected humans to have to assemble and manage groups of AI agents carefully. Newer systems, he argues, increasingly plan, delegate, and divide roles on their own.

Thousands of agents handled millions of handoffs without a human dispatcher

Mollick builds the argument from several observations. With personal agents such as Muse and Dots, the important change is how much context gathering, planning, and follow-up no longer falls on the user. An agent watches connected accounts, prepares next steps, and reaches out when it encounters a problem or needs approval.

His largest example concerns an OpenAI effort around the Navier-Stokes existence and smoothness problem. According to the account, thousands of agents worked for 88 hours and exchanged about 2.7 million messages. The result had not received formal acceptance from the Clay Mathematics Institute when Mollick published, so this remains a claim and interpretation from participants rather than a settled mathematical verdict.

Management value shifts from assigning work to defining the outcome

For product and engineering teams, the organizational consequence matters more than the dramatic number. When a model can create helper agents, distribute research among them, and redirect effort as work unfolds, a hand-built multi-agent workflow ages quickly. A well-specified objective, a verifiable output, and firm boundaries are more durable assets.

Research from SwarmWorld points in the same direction. In populations of 50 to 200 initially identical agents, explorer, builder, and maintenance roles emerged without prior assignment. Yet the shared environment did not win on every metric. Independent search remained competitive for producing the strongest single artifact.

Self-organization lowers coordination costs, not accountability

Agents being able to coordinate does not mean they can choose the right objective. A system can move efficiently in the wrong direction, propagate a false assumption, or use a shared environment in ways that ordinary conversation monitoring misses. The less a human directs individual steps, the more important sandboxes, budgets, permissions, and independent output checks become.

Mollick also joins personal agents, research experiments, and exceptional incidents in one argument. That is a useful picture of a trend, not evidence that an autonomous team can reliably perform routine corporate work for weeks without supervision.

Verifiable outcomes outside the swarm will decide whether it pays off

The first useful signal will not be the number of agents launched. It will be the share of tasks for which a team defines a test, a cost ceiling, and a human review checkpoint in advance, then gets a better result than a simpler single-agent setup.

Companies should also monitor traces in shared environments: changes to files, databases, queues, and permissions. If agents can coordinate through artifacts rather than messages, chat audits reveal only part of the organization that actually emerged.

Lilith's verdict

When the swarm assigns its own shifts, the human is no longer the manager at the whiteboard. The human locks the dangerous doors and decides what counts as done.