Ethan Mollick, a professor at Wharton and an influential AI observer, proposes applying lessons from organizational theory to agentic AI systems. His argument: decades of research on human organizations offer frameworks directly applicable to the challenges of multi-agent coordination.
The spans of control problem: Current agentic systems implicitly assume that models can manage an unlimited number of sub-agents, which is clearly false. A human manager tops out at fewer than ten effective direct reports. Mollick estimates that a hundred sub-agents far exceeds the capacity of an orchestrator agent. His provocative solution: create "middle management agents" - an intermediate hierarchy between the main orchestrator and the execution agents.
Boundary objects: In organizational theory, boundary objects are artifacts passed between groups (marketing, IT, sales) to convey meaning when a project crosses boundaries - prototypes, user stories, specifications. Currently, AI agents exchange plain text and sometimes code. Mollick advocates for structured boundary objects that agents of different capability levels can read and modify. This approach would resolve many coordination failures while reducing token consumption.
The coupling problem: Coupling measures the degree of connection between organizational units. Most agentic systems are either too tightly coupled (every step requires human approval) or too loosely coupled (loss of control and coherence). This tradeoff is well studied in organizational theory, and Mollick bets that many findings apply directly to agent architectures.
Bounded rationality: A foundational concept in organizational science, bounded rationality describes how decision-makers operate with incomplete information and finite cognitive capacity. This framework likely applies to AI agents facing large contexts and complex decisions.
Critique of the labs: Mollick observes that everyone is rushing toward "agent swarms" (a term he calls "terribly named") without realizing that success will not depend solely on model quality, but on organizational design choices. He doubts that AI labs perceive this dimension and calls for more experimentation led by people who understand real human coordination problems.