Capability is not authority. An agent’s ability to organize work does not establish its right to perform every action that work makes possible. Better coordination changes how execution is arranged; it does not settle which outcome deserves pursuit, who may authorize its consequences, or what conditions must hold before those consequences are permitted.
That distinction matters precisely when the agent becomes more capable. An inability to act can conceal an unresolved permission boundary. Once the system can act, the boundary has to stand on something other than the system’s limitations. The relevant question is no longer only whether the work can be done. It is whether this actor may do this work, for this purpose, under these conditions.
The Dot and the Swarm—a post attributed to Ethan Mollick in a forwarded PDF with unauthenticated sender and authorship—describes agents coordinating with little human management while people set goals and redirected effort. In that description, humans did not specify every exchange among agents. They selected problems, changed the direction of effort, and reassessed where the work should go.
The useful distinction is between organizing the work and deciding where to point it. If that division holds, the orchestration problem changes. It does not disappear. Less human involvement in arranging execution can coexist with a continuing need to govern the interface between execution and consequential decisions. The freedom to find a route is different from the authority to choose any destination.
The serious counterargument is that elaborate orchestration can become obsolete scaffolding. The post applies the Bitter Lesson to organizational design: functions once assigned to carefully constructed information systems, prompt chains, and management structures may increasingly be performed by models themselves. A system that insists on prescribing every step can preserve a human diagram after the diagram has stopped contributing to the result.
That objection deserves more than a concession in passing. If the model can competently discover the relevant information, arrange its subtasks, and revise its plan, requiring a person to specify those same operations may add delay without adding judgment. A handoff is not valuable merely because someone designed it. A fixed sequence is not safer merely because it is visible. Its value depends on the condition it actually satisfies.
The answer is to loosen choreography without confusing that freedom with permission. An agent can be allowed to choose how it performs authorized work without receiving authority over every adjacent action. A mandate can leave substantial room for adaptation while retaining limits on purpose, access, expenditure, or external commitments. The important distinction is not between freedom and control in the abstract. It is between freedom inside a legitimate scope and an expansion of that scope.
There is an equally serious objection from the other direction: governance can become the bottleneck it claims to prevent. If every harmless rearrangement requires a fresh decision, little remains of the advantage of self-organization. The design question is therefore not how to maximize the number of approval points. It is how to preserve consequential boundaries without turning all execution into a succession of avoidable waits.
This becomes clearer once a dependency is understood as more than another task. A dependency may be missing evidence. It may be a decision that only a particular person is entitled to make. It may be permission required before an external action. These conditions can affect the same workflow, but they are not interchangeable. More task coordination does not necessarily resolve any of them.
Consider a hypothetical purchasing workflow. An agent might compare suppliers, identify unanswered questions, and assemble a recommendation. Another task could improve that recommendation. But the remaining dependency might instead be a budget decision or authority to place an order. Finding the best supplier would not, by itself, resolve either condition. The analysis could be excellent while the purchase remained unauthorized.
The distinction is not peculiar to large transactions. The Dot and the Swarm describes an agent noticing an error in a permit email and preparing a correction. It separately describes an agent contacting an airline to request an extension after being asked.
Those examples illustrate a useful difference: identifying an issue, preparing a response, and acting externally are different kinds of movement. A prepared correction is not a sent correction. A recommendation is not a decision. A decision is not necessarily an accepted commitment, and an attempted action is not necessarily a completed one. Treating these as interchangeable makes activity appear to settle a condition that remains open. Activity is not an accepted commitment: progress has to be described in terms of the state that actually changed.
The highest-value next move is therefore not automatically the one that produces the most activity, or even the one that removes a dependency. Its value depends on the remaining value of the outcome, the cost of reaching it, and whether it can still be reached in time. A useful criterion is to identify the dependency blocking an authorized outcome and ask what would resolve it, then test whether resolving it is still worth the required effort before the deadline.
What must be true before the next consequential action? Who can make that condition true? Who is authorized to decide whether the resulting action should proceed? What evidence would show that the dependency has been resolved? The dependency resolver and the authorized decision-maker may be the same actor, but they need not be. Someone can supply missing information without having the right to make the decision that information enables. These questions connect the work to its purpose rather than to the size of its task list.
That does not mean every agent should stop whenever any uncertainty appears. Work already within scope can continue when its expected contribution still justifies the remaining cost, it remains within the cost ceiling and deadline, and the relevant outcome checks still support proceeding. Research may narrow a decision. A comparison may reveal that an apparent choice is not a meaningful one. An agent may discover a different route that avoids an unnecessary dependency altogether. The boundary should be retained where a new decision or authority is genuinely required, not imposed indiscriminately on all useful preparation.
Sometimes the missing condition is information, and further investigation is the right move. Sometimes the information is already sufficient, and another research pass merely delays the person who must decide. Sometimes the decision has been made, but the action still lacks the required permission or capability. The practical question is which condition is still false, and what available work can change it at a cost and in time that still serve the outcome. Resolving a dependency makes a route available; it does not automatically justify taking it.
The same discipline has to continue after an agent acts. There are two separate problems. An action can occur without legitimate permission. A report can describe an action inaccurately. Improving coordination among agents does not, by itself, establish either authorized conduct or a reliable report of the result. A system can become more effective at pursuing a course of action while leaving its relationship to the principal unresolved.
For consequential work, the check should match the claim. Evidence that a response was prepared supports preparation. Evidence that an external system accepted a request supports acceptance of that request. Evidence of the resulting change supports the corresponding outcome. Those may be closely connected, but proximity is not identity. The report of what happened should not leap over the condition that the evidence has yet to establish.
None of this is an argument for governance in place of capability. The post explicitly says AI remains limited and leaves uncertain how well self-organizing agents handle the long, unglamorous work that fills much of an organization’s time. The ability to coordinate an impressive effort does not settle reliability across every kind of work an organization needs.
Permission cannot compensate for inability. Nor does effective coordination substitute for accountability. An authorized system that repeatedly fails to perform the task is not adequate merely because its boundaries are clear. A capable system that exceeds its mandate is not adequate merely because its output is useful. Both constraints belong in the operating design.
This also limits what can reasonably be concluded from either side of the argument. Replacing a brittle task sequence with adaptive agents may improve execution without resolving authority. Adding approval gates may constrain action without establishing competence. The relevant conditions have to be considered together; success on one is not evidence that the other has been satisfied.
Orchestration earns its place by making consequential dependencies legible and helping resolve them. When it merely dictates steps the model can arrange for itself, it risks becoming the bottleneck it was meant to remove. When it exposes the difference between missing information, an unresolved decision, absent permission, and unconfirmed completion, it serves a purpose that better self-organization does not automatically erase.
The aim is not to keep humans arranging every exchange. It is to ensure that increasingly capable execution remains connected to a legitimate purpose, an appropriate scope, and evidence of what actually happened. Let the agent discover an effective route. Do not let discovering the route establish its permission to cross every boundary along it.
As always, context is all.
Source note
Based on personal communication with Breyden Taylor and Sabeel Ahmed, retrieved October 4, 2026.