From Economic Agents to Agentic Economies: A System... | AI Research

From Economic Agents to Agentic Economies: A Systems Blueprint for Economic World Models proposes a framework for building generative simulations of economies. The authors aim to move beyond traditional economic modeling—which often focuses on observing and predicting aggregate outcomes—toward building "world models" that simulate the internal dynamics of an economy, including how heterogeneous agents interact, adapt, and co-evolve with market institutions.

The Concept of Economic World Models

The authors define an Economic World Model (EWM) as a computational system that simulates how an economy evolves from within. Unlike physical world models, where state transitions are governed by external, objective laws, economic worlds are driven by the subjective beliefs and strategic decisions of agents. In an EWM, the state of the economy at any given time is a combination of aggregate variables (such as prices and output), the private states of individual agents (such as wealth and balance sheets), their subjective beliefs, and the institutional rules governing their interactions.

Implementation as a Modular Runtime

The paper outlines an implementation roadmap that treats an EWM as a modular software runtime rather than a monolithic predictor. This runtime operates through a continuous feedback loop: 1. Agent Layer: Heterogeneous actors observe the current state and choose actions based on their objectives and beliefs. 2. Environment Layer: Market mechanisms and institutional rules process these actions to determine the next aggregate state. 3. Co-evolution Layer: Agents and institutions adapt their strategies and rules over time. 4. Alignment Layer: The system compares simulated outcomes with real-world empirical data to correct model drift.
The authors provide a minimal execution interface, suggesting that developers can construct these worlds by defining agent populations (e.g., households, firms, banks) and stepping through the simulation to observe how individual actions generate aggregate economic outcomes.

Engineering Requirements

To ensure these models are functional and reliable, the authors establish four engineering desiderata:

• Endogenous Closure: Key economic outcomes, such as prices and allocations, must emerge from agent interactions rather than being imposed by the system.

• Behavioral Fidelity: Agents must reflect the complexity of real-world actors, including bounded rationality and context-dependent adaptation, rather than relying on fixed rules.

• Evolving Dynamics: The system must account for the co-evolution of agent strategies and institutional rules.

• Reality Alignment: The model must maintain a connection to the real economy through persistent empirical validation.

Current Limitations and Future Directions

The authors conducted a systematic literature survey and concluded that existing research is largely concentrated in lower-level simulation environments. Systems that incorporate self-evolving agents, endogenous institutions, and persistent alignment with real-world data remain rare. The paper suggests that the rise of Large Language Models (LLMs) and agentic AI provides the necessary technical substrate to overcome these gaps, offering a path toward creating high-fidelity sandboxes for policy testing, risk assessment, and AI agent training.