Specialized decision-makers plus many simple agents can outperform a society of uniformly informed individuals: when information is costly, delegation and cognitive division of labour preserve much of knowledge's value while cutting decision costs.

Main Finding

When information is valuable but costly to acquire, process, and coordinate, evolution (and selection more generally) can favor heterogeneous cognitive organization: most agents remain simple, relying on low-cost heuristics, while a minority (individuals, roles, or institutions) specialize in costly information processing. This cognitive division of labour — “decision-compression” institutions that translate complex states into simple actionable signals — can dominate universal individual complexity whenever the cognitive cost savings from distributed simplicity exceed the utility lost to imperfect delegation. The specialist need not be a volunteer: the role can be sustained by private returns (rents, status, control, superior information).

Key Points

• Puzzle addressed: If information improves decisions, why do populations not evolve into universally complex/informed agents?
• Core mechanism: Information has positive utility but non-negligible acquisition/processing/coordination costs (metabolic, attention, computational, institutional).
• Three social configurations compared:
  • Uniformly complex agents (everyone acquires/processes much information).
  • Uniformly simple agents (everyone uses low-cost heuristics).
  • Heterogeneous population: simple agents plus a centralized/specialized decision-maker (decision-compression institution).
• Trade-offs captured:
  • Benefits of extra information (improved decision quality).
  • Costs of cognition and coordination.
  • Delegation losses when simple agents follow outputs of specialists (imperfect fit, reduced autonomy).
  • Private returns to the specialist (preventing volunteer’s dilemma).
• Main theoretical prediction: Heterogeneous organization with specialized decision-processing dominates when cost savings from widespread simplicity exceed delegation/agency losses. Decision-compression institutions can preserve much of the value of information at lower collective cost.
• Broader connections: links bounded rationality, rational inattention, hierarchy/firm theory, market price-information aggregation, cultural evolution, and division of labour.
• Conceptual reframing: Simplicity is adaptive and a precondition for scalable social organization, not merely a failure or limitation.

Data & Methods

• Nature of the study: theoretical and computational/modeling (no primary empirical dataset reported).
• Formal model components:
  • Agents face decision problems in complex environments with uncertain states.
  • Utility decomposition includes: payoff from chosen action given state (improved by information), costs of acquiring/processing information (cognitive/metabolic/attention), coordination/communication costs, and delegation losses when following a specialist.
  • Population structures formally defined: uniform-complexity society, uniform-simplicity society, heterogeneous society with a decision-making centre.
  • Institutional rents/benefits can accrue to specialists (status, control, information advantage) and are modeled as private payoff terms.
• Analysis:
  • Analytical comparison of expected utilities across the three configurations, deriving conditions where heterogeneous organization is preferred.
  • Numerical exploration/simulations across parameter space (environmental complexity, cognitive cost parameters, delegation accuracy, size of private returns) to illustrate regions where each configuration dominates.
  • Qualitative discussion linking model results to existing literatures (rational inattention, bounded rationality, theory of the firm, division of labour, evolutionary cost of cognition).
• Limitations of methods: model is stylized; delegation losses, institutional forms, and the nature of rents are abstracted rather than empirically parametrized.

Implications for AI Economics

• Division of cognitive labour maps directly to AI ecosystems:
  • Specialized AI systems/institutions (platforms, orchestration agents, expert models) can function as decision-compression centres that process high-dimensional information and emit simplified signals or policies for human or simpler agents.
  • Widespread deployment of simple-user-facing agents plus specialized backend AI may be more efficient than making every actor individually highly complex (expensive compute, latency, energy).
• Cost-benefit trade-offs in AI design:
  • Compute, attention, latency, and energy costs play the role of biological/informational costs—these can make centralized or specialized AI processing socially optimal.
  • Delegation losses correspond to misalignment, loss of context, reduced autonomy, and trust costs when humans follow AI outputs. These impose a limit on how far delegation should extend before centralized processing becomes net harmful.
• Market and institutional implications:
  • Providers of specialized AI can capture rents (platform fees, information advantages, reputational/status-based control), which can sustain the specialist role without requiring altruism. That explains persistent centralization in some markets.
  • Competition, transparency, and regulation will shape how rents are distributed and therefore the incentive to provide specialist decision services.
• Policy and design considerations:
  • Encourage mechanisms that reduce delegation losses: interpretability, feedback loops, improved context capture, local customization, accountability, and incentives aligned with principals.
  • Monitor concentration risks: if private returns to specialists are large and persistent, decision-making power may centralize excessively, raising systemic risk and welfare concerns.
  • Invest in coordination technologies and standards (data interchange, protocols) that lower institutional/coordination costs, potentially shifting the optimal balance between specialization and distributed complexity.
• Predictions for AI-driven economies:
  • Heterogeneous capability distributions (few large, sophisticated models + many lightweight assistants) are likely efficient in many settings.
  • In environments where delegation is highly reliable and delegation losses are small, centralization increases; where delegation loss or trust costs are high, more agents will retain local complexity.
  • Technological shifts that reduce the cost of individual information processing (cheap compute, low-latency personal models, better local sensors) will move optima toward more distributed complexity; conversely, large fixed-cost gains to scale (very expensive models, network effects) favor centralized specialists.
• Research directions:
  • Empirical calibration: measure delegation loss, cognitive/compute costs, and private returns to specialists in real AI markets.
  • Mechanism design for robust, incentive-aligned delegation (contracts, verification, reputation).
  • Study dynamics of transition regimes as compute/communication costs change (path dependency, lock-in).
  • Explore welfare trade-offs from centralization in safety, fairness, and market competition contexts.

Short summary takeaway: The paper formalizes why and when societies (biological or socioeconomic) evolve to have many simple agents and a few information specialists. For AI economics, the framework explains the prevalence and persistence of specialized AI services and gives a lens for evaluating trade-offs between centralized, sophisticated AI providers and distributed, local intelligence — emphasizing costs of information processing, delegation losses, and the role of rents/incentives in sustaining specialization.