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 Cong
cs.AI, cs.LG
2026-08-06
A 737-paper survey ranks Economic World Models on a six-level capability ladder. Agents keep getting smarter; evolving institutions and real-world alignment remain nearly absent.
Economics advances on an observe-then-explain footing: economists record aggregate variables, gather micro evidence, estimate causal effects, then ask what caused what. That can predict a turning point, but it leaves a deeper question open. Can a phenomenon be made to emerge from inside the model? Deriving an equilibrium or fitting parameters does not by itself explain how the outcome came about.
This paper wants to move economics from retrospective explanation to generative simulation. The Economic World Model (EWM) it lays out is a generative engine: heterogeneous actors (households, firms, banks, regulators), their beliefs and actions, and the markets and institutions around them are placed in a computable environment, and macro-level states emerge from their interaction. The backdrop is AI's push to extend world models beyond physical environments, together with the fact that OpenAI appointed its first chief economist in 2024 and Anthropic launched an Economic Index, signs that industry wants to understand AI's economic consequences. The question is where this effort stands and what is missing.
The authors formalize an EWM as a transition operator, s{t+1} = T(st, ut): the current economic state plus an external intervention yields the next state. What separates it from a physical world model is who drives the transition. In a physical world, motion follows external laws that do not depend on what an object believes. In an economic world, the next state is produced endogenously by agents that form beliefs, strategize, and change their minds.
Around that definition they propose a four-layer modular runtime: an agent layer (objectives, information, constraints, memory, feasible actions), an environment layer (economic states, market mechanisms, contracts, accounting identities, institutional rules), a co-evolution layer (agents, strategies, mechanisms, and institutions adapt over repeated rollouts), and a real-world alignment layer that compares simulated trajectories with empirical observation and corrects drift. The target is an agentic economy: agents train inside the EWM, grow more capable, and are written back into the EWM's own agents, sharpening next-state prediction in a self-reinforcing loop.
The load-bearing contribution is a six-level capability ladder that grades implementation maturity:
| Level | Name | Marker |
| L1 | Fixed rule-based agents | Agents and environment follow preset rules, no online adaptation |
| L2 | Adaptive rule-based agents | Agents update policies via RL, evolution, or online learning, but hold no rich beliefs |
| L3 | LLM-based autonomous agents | Agents carry beliefs, memory, language, reasoning |
| L4 | Self-evolving agents | Agents persistently acquire new strategies, skills, tools |
| L5 | Agent-environment co-evolution | Institutions, rules, governance change endogenously with outcomes |
| L6 | Sim-to-real economic twins | Online correction of simulated trajectories against real observation |
The ladder measures implementation capability only; it does not by itself guarantee counterfactual consistency in the economic sense, which is the job of the DDGE (data-driven generative equilibrium) framework from Cong (2025).
To locate the field on this ladder, the authors ran a systematic survey: searching eight arXiv categories and the UTD-24 business journal list, merging and deduplicating to 7,836 candidates (6,008 arXiv plus 1,828 UTD-24), then a two-stage LLM screen (GPT-5.4-mini on abstracts, GPT-5.5 on full text), with borderline cases flagged for manual review. That yields 737 validated EWM papers.
The survey runs no experiments; its results are a portrait of those 737 papers, and the picture is clear.
Agents are the frontier of recent progress. Volume grows steadily after 2018 and jumps in 2024 and 2025, almost entirely in L1 to L3, especially LLM-based autonomous agents that learn and reason. Business journals sit at the bottom: the UTD-24 panel clusters almost entirely in L1 and L2, with strong economic discipline and clear institutional structure, but almost no LLM agents, no persistent self-evolution, and no real-world alignment. The split is a division of labor, with arXiv pushing new AI capability and business journals supplying more grounded economic modeling.
The top is nearly empty. L4 to L6 are a thin sliver of the landscape. L5, with endogenous institutional change, is rare, meaning most simulations still treat the economic environment as a fixed stage. L6 sim-to-real twins are close to absent, with only scattered early prototypes that correct against real data. The field has made agents smarter without solving the harder systems problem: putting self-evolving agents, evolving institutions, and online real-world alignment inside one economic world.
For practitioners, the ladder is a map: it shows where you can build and where the open ground is. If you build agent training environments, policy sandboxes, strategy rehearsal, or AI safety testbeds (to detect manipulation, collusion, cascading failure), the upper levels are where the value and the gap both are. The applications are concrete: for humans, a high-fidelity sandbox; for AI agents, a substrate for training, planning, evaluation, and safety.
One honest flag: this is a survey plus a framework proposal. The authors do not build an L6 system, and there is no empirical evaluation. Its value is in defining the problem and supplying a taxonomy and interface starting point, not in shipping a working system. Read it as a roadmap, not a product.
The deepest one first: it is a framework proposal with no empirical validation. The six-level ladder, the four-layer runtime, and the co-evolution loop are design-level claims; no end-to-end high-fidelity economic world is built to show they work.
Classification is LLM-assisted (GPT-5.5 on full text) and carries subjectivity, with borderlines checked by hand, so the 737-paper total and the level shares will shift with the grading rubric. The authors name their own open challenges: behavioral realism (aligning to real behavioral data), economic closure (nonlinear aggregation that stays both feasible and scalable), co-evolution (what evolves, when, and how to stop drift), computational scalability, and world-level evaluation, where counterfactual worlds have no single ground-truth trajectory.
It is built on Cong (2025)'s DDGE, itself a working paper whose counterfactual-consistency theory is unvalidated. The most worthwhile gaps to chase are L5 and L6, where institutions and real data enter the loop.