From Economic Agents to Agentic Economies: A Systems Blueprint for Economic World Models

Jiale Han, Xiang Li, Jing Qian, Wenyuan Gu, Pin Gao, Ye Luo, Hongyuan Zha, Dacheng Tao, Benyou Wang, Lin William CongarXiv 2026 · arXiv

TL;DR

This paper introduces a computer science and AI systems blueprint for Economic World Models (EWMs)—generative simulation engines where heterogeneous agents act, interact, and co-evolve with market mechanisms and institutions. By establishing a six-level capability ladder and conducting a massive literature survey of 737 validated papers, the authors map out a clear engineering roadmap to transition from isolated economic AI agents to fully realized, real-world-aligned "Agentic Economies."

What Problem Does This Paper Address?

Traditional economics relies heavily on an "observe and explain" paradigm, using econometric models or highly stylized mathematical frameworks (like DSGE or CGE models) to forecast outcomes. However, these traditional approaches struggle with generative explanation—they cannot easily simulate how complex macroeconomic phenomena organically emerge from individual, micro-level interactions, especially under bounded rationality and qualitative, narrative-driven market shifts.

Conversely, standard AI "world models" excel at predicting physical environments (e.g., autonomous driving or robotics). But as Isaac Newton famously lamented, calculating the madness of people is far harder than calculating the motion of heavenly bodies. Physical worlds operate under stable, objective, and exogenous laws. Economic worlds are fundamentally different: they are driven by subjective agents whose beliefs, expectations, and strategic actions endogenously alter the very state of the world, creating a highly complex, reflexive feedback loop.

Figure 1: Physical versus economic worlds. Physical state transitions are governed by external laws, whereas economic state transitions are generated by heterogeneous agents interacting through markets, institutions, and evolving rules.

This paper bridges this gap by providing a formal systems architecture to make economic environments buildable, testable, and alignable as interactive AI sandboxes.

Methods

The authors formalize the EWM framework by defining the state of an economic world at time $t$ as a multi-part tuple: $$s_t = (x_t, \{z_t^i\}_{i=1}^{N_t}, \{b_t^i\}_{i=1}^{N_t}, I_t)$$ Where $x_t$ represents the aggregate economic state (prices, volatility), $z_t^i$ is the private state of an agent (cash, inventory), $b_t^i$ is the agent's subjective belief state, and $I_t$ represents the institutional environment (rules, tax schedules).

The core runtime architecture decomposes state transitions into an engineered closed loop across four distinct layers:

  1. Agent Layer: Instantiates heterogeneous actors with specific objectives, memories, and tools.
  2. Environment Layer: Encodes market clearing mechanisms, contracts, and hard institutional constraints.
  3. Co-evolution Layer: Governs how agent strategies and institutional rules adapt dynamically over repeated rollouts.
  4. Real-world Alignment Layer: Continuously anchors the simulation to empirical data to prevent model drift.

To categorize the maturity of these systems, the authors propose a Six-Level Capability Ladder:

  • L1 (Fixed Rule-Based): Pre-specified rules for both agents and environments.
  • L2 (Adaptive Rule-Based): Agents adapt using symbolic algorithms (e.g., Reinforcement Learning).
  • L3 (LLM-Based Autonomous): Agents reason using language, explicit beliefs, and memory.
  • L4 (Self-Evolving Agents): Agents persistently expand their skills and strategy libraries.
  • L5 (Evolving Economic Worlds): Institutional rules and market mechanisms adapt endogenously.
  • L6 (Sim-to-Real Economic Twins): The running simulation continuously ingests live, real-world data streams for real-time correction.

Datasets

Rather than introducing a new empirical dataset, the paper presents a comprehensive, systematic literature survey. The authors built an LLM-assisted screening pipeline (utilizing GPT-5.4-mini and GPT-5.5) to filter a candidate pool of 7,836 papers published between 1950 and April 2026 across two primary repositories:

  • arXiv: Spanning categories in economics, finance, multi-agent systems, and machine learning.
  • Web of Science (UTD-24 journal list): Covering peer-reviewed research in management, finance, and economics.

The pipeline successfully validated and classified 737 papers as true Economic World Models based on strict engineering desiderata.

Key Insights and Ideas

  • The AI vs. Econ Division of Labor: The literature survey revealed a stark divide. Preprints on arXiv drive rapid progress in agent intelligence (L3/L4) using LLMs, but often treat the environment as a passive backdrop. Conversely, top business and economics journals (UTD-24) excel at institutional discipline and structural realism (L1/L2) but rarely incorporate advanced cognitive AI architectures.
  • The Level 6 Gap: True "Sim-to-Real Economic Twins" (Level 6) are exceptionally rare. Most existing attempts at empirical alignment happen offline during the model-fitting stage rather than acting as a live, continuous correction signal within a running simulation loop.
  • Co-Improvement Cycle: The ultimate value of an EWM lies in a bidirectional training loop: capable agents expose flaws and edge cases in the simulated market mechanisms, while a high-fidelity market environment provides a rigorous substrate to train safer, more robust AI agents.

Key Results

The systematic survey mapped out the historical trajectory of EWM research, highlighting a massive explosion of interest starting around 2024, driven primarily by large language models. However, the distribution across capability levels is heavily skewed:

  • The vast majority of historical and recent work remains tightly clustered in Levels 1, 2, and 3.
  • Systems featuring self-evolving agents (L4) or endogenous institutional evolution (L5) represent only a tiny fraction of the literature.
  • Fully realized Level 6 systems remain the most critical, unaddressed frontier in the field.

Limitations and Future Work

The authors identify several open challenges that must be solved to realize true agentic economies:

  • Behavioral Realism: Ensuring that LLM-based economic agents accurately replicate human biases, risk aversion, and decision-making under scarcity, rather than just generating plausible text.
  • Non-linear Economic Closure: Scaling up market-clearing mechanisms computationally without breaking stock-flow consistency or oversimplifying strategic interactions.
  • Co-evolutionary Stability: Preventing self-evolving agents and shifting institutional rules from drifting into highly unrealistic or completely unstable simulated dynamics.
  • Equilibrium Discipline: Integrating systems architecture with rigorous economic counterfactual consistency frameworks, such as Data-Driven Generative Equilibrium (DDGE).

Summary updated

Abstract

Economic World Models (EWMs) are generative economic models that simulate how economies evolve from within by modeling heterogeneous agents, their beliefs and actions, and the market and institutional mechanisms through which their interactions produce aggregate outcomes. This paper develops an implementation roadmap for building economic world models as generative engines in which heterogeneous agents act, interact, adapt, and co-evolve with markets and institutions, thereby producing economic dynamics from the inside. We organize EWM systems into a six-level capability ladder, from fixed rule-based agent worlds to adaptive and LLM-based agent worlds, self-evolving agents, evolving institutional worlds, and sim-to-real economic twins aligned with real observations. A systematic literature survey across these levels reveals that existing work remains concentrated in lower-level agent and simulation environments, while systems with self-evolving agents, endogenous institutions, persistent empirical alignment, and validated economic mechanisms remain rare. By translating the EWM agenda into an implementation blueprint, this paper aims to accelerate the development of the next generation of economic simulation environments that can serve as high-fidelity sandboxes for human decision-makers and as training, planning, evaluation, and safety substrates for AI agents. We release a curated paper list and related resources to support future research.
BibTeX
@misc{han2026from,
  author = {Jiale Han and Xiang Li and Jing Qian and Wenyuan Gu and Pin Gao and Ye Luo and Hongyuan Zha and Dacheng Tao and Benyou Wang and Lin William Cong},
  title = {{From Economic Agents to Agentic Economies: A Systems Blueprint for Economic World Models}},
  year = {2026},
  eprint = {2608.06020},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2608.06020},
}
Figure