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

This paper develops a CS/AI systems implementation roadmap for building Economic World Models (EWMs) — 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. The authors frame their work as complementary to the economic framework of Cong (2025), which introduces the EWM/DDGE (Data-Driven Generative Equilibrium) framework. Whereas Cong's work takes an economics perspective, "the present paper takes a CS/AI systems perspective: it asks how such economic worlds can be implemented, evaluated, and scaled as computational environments for agent training, policy sandboxes, planning engines, and sim-to-real economic twins."

The paper's working definition states: "An Economic World Model system is a generative engine for economic environments: it predicts how an economy moves by modeling how the agents inside observe, reason, act, interact, and adapt under economic institutions and constraints."

The central distinction from physical world models is that physical state transitions are governed by external laws, whereas economic state transitions are generated by heterogeneous agents interacting through markets, institutions, and evolving rules. The paper quotes Newton — I can calculate the motion of heavenly bodies, but not the madness of people — to illustrate that economic worlds are fundamentally different: unlike planets, economic actors interpret information, form beliefs, and act strategically.

Four features make economic worlds especially distinct: subjective agents (who form beliefs and expectations under incomplete information), endogenous outcomes (prices, allocations, and macro conditions emerge from decentralized interactions rather than being externally imposed), reflexive feedback (economic feedback is belief-mediated and self-referential—agents act on beliefs, their actions change the economy, which reshapes later beliefs), and adaptive dynamics (agents learn and institutions evolve over time).

The paper distinguishes between a stand-alone Economic Agent and a full Agentic Economy: "A stand-alone Economic Agent primarily performs agent-side next-action generation... What it lacks is an executable economic feedback loop in which the consequences of that action are realized, evaluated, and returned as learning signals. An Agentic Economy, by contrast, turns this one-sided prediction problem into an engineered closed loop between agent-side next-action prediction and world-side next-state prediction."

The paper describes an iterative co-improvement cycle between agents and the EWM: agents are trained through simulated rollouts within the EWM; as agents become more capable or behaviorally accurate, these advances transfer back to agents inside the EWM; this improves next-action prediction, which leads to more accurate next-state prediction; the improved EWM then provides a better training environment for subsequent agent training — creating a co-improvement cycle between economic agents and the EWM.

The paper formally defines an economic state as a typed snapshot containing three parts:

• Aggregate economic state x t: economic variables and information such as prices, quantities, volatility, employment, output, interest rates, public news and other economy-wide conditions

• Agent-side states z t i, b t i i=1 N t: private states z t i (cash, wealth, inventory, balance-sheet position, contracts, production capacity, risk exposure, outstanding obligations) and belief states b t i (expectations, perceived risks, forecasts, uncertainty, and beliefs about prices, demand, policies, or other agents' behavior)

• Institutional environment I t: the rules under which agents act, including constraints, market mechanisms, policy regimes, disclosure rules, information channels, and data-access permissions

The state transition decomposes into five operators: belief update B i, next-action prediction pi i, agent state update i, endogenous aggregation G, and rules/mechanisms evolution.

The paper specifies four engineering desiderata for strong EWM systems:

1. Endogenous closure: Prices, allocations, and macroeconomic conditions should emerge from decentralized interactions among heterogeneous agents, rather than be imposed from outside.

2. Behavioral fidelity: "Agent behavior should reflect how households, firms, banks, and investors perceive, interpret, and respond to incentives, information, and constraints... What matters is capturing heterogeneity, bounded rationality, and context-dependent adaptation. The paper notes that Recent evidence shows that LLMs can exhibit systematic behavioral patterns in economic and financial decision tasks... EWM agents therefore require behavioral validation rather than persona prompts alone."

3. Evolving dynamics: both the agents and the governing rules of the economy should co-evolve over time. The model should represent the evolution of strategies, cognition, and governing rules at each of these layers.

4. Reality alignment: The EWM should not only generate plausible simulations, but also repeatedly compare them with newly observed outcomes and update its internal dynamics accordingly — progressively narrowing the gap between the simulated economy and the real one.

The paper organizes EWM systems into a six-level capability ladder:

• L1: Fixed rule-based agent worlds — agents and the environment follow fixed, pre-specified rules. The cognitive operator B i is inactive and institutional evolution is inactive. Provides endogenous closure but no in-world adaptation.

• L2: Adaptive rule-based agent worlds — agents revise their decision policies through symbolic or algorithmic mechanisms such as reinforcement learning, evolutionary search, online learning, or adaptive heuristics. Still no rich beliefs (cognitive operator remains inactive).

• L3: LLM-based autonomous agent worlds — agents are equipped with explicit cognitive states, including beliefs, expectations, memory, and subjective representations of the economy. The cognitive operator becomes active; agents reason, communicate, and make context-sensitive economic decisions. The cognitive substrate stays fixed during rollout.

• L4: Self-evolving agent worlds — adaptation extends from beliefs and actions to the agent capability substrate itself. Agents may acquire new strategies, skills, tools, memories, or behavioral routines through in-world experience. The institutional environment remains fixed.

• L5: Evolving economic worlds — the institutional evolution operator becomes active, allowing policies, market rules, contracts, governance structures, or institutional constraints to change in response to agent behavior and aggregate outcomes. What defines this level is whether the rules of the game become endogenous.

• L6: Sim-to-real economic twins — The EWM maintains an online correction loop that compares simulated trajectories with newly observed economic data during the rollout. This creates a two-loop architecture: The inner loop generates endogenous economic dynamics, while the outer loop performs real-time monitoring, diagnosis, and correction.

The paper emphasizes that this hierarchy is "an implementation taxonomy rather than an equilibrium concept: a system can be high on the capability ladder and still fail counterfactual consistency if learned components are not re-centered on the data generated by the counterfactual itself, the problem addressed by DDGE in Cong (2025)."

The paper conducted a systematic literature collection and classification. Stage one used intersection-based keyword filtering across arXiv (eight categories: econ.GN, econ.TH, econ.EM, cs.MA, cs.AI, cs.LG, stat.AP, q-fin.ST) and Web of Science (UTD-24 journal list), covering 1950–April 2026, yielding a pool of 7,836 candidate papers. Stage two used GPT-5.4-mini for recall-oriented title-and-abstract screening, reducing the pool to 794 arXiv papers and 87 UTD-24 papers. Stage three used GPT-5.5 for precision-oriented full-text classification, with borderline cases flagged for manual author verification. This yielded a final corpus of 737 validated EWM papers.

Key findings from the literature landscape:

• Agent intelligence has become the main frontier of recent progress. The arXiv panel shows acceleration after 2018, with a sharp rise around 2024–2025, concentrated in L1–L3 systems. Most of this growth is concentrated in L1–L3, especially systems with adaptive rule-based agents and LLM-based autonomous agents.

• Business and economics journals remain concentrated at lower levels. UTD-24 papers are concentrated almost entirely in L1 and L2, reflecting the longer tradition of economics, finance, management, and operations research in building rule-based or adaptive agent-based models with explicit mechanisms and economic interpretation. The contrast reveals a division of labor: arXiv pushes new AI capabilities, while business and economics journals provide more economically grounded modeling traditions.

• Higher-level EWMs remain rare. L4 systems are recent and uncommon; L5 systems with endogenous institutional or rule evolution are even rarer, indicating that most simulations still treat the economic environment as a fixed stage rather than an evolving object; L6 systems are nearly absent, with only very limited evidence of repeated correction against real-world observations.

Representative L6 works are notably sparse: only Wiesinger et al. (2010), reverse-engineering traders from real Nasdaq data, and Evans et al. (2025), which formalize adaptive ABM as a Stackelberg game against experimental data. The paper notes these fall short of the level's full ambition. Empirical feedback in both cases enters the system at the model-fitting stage... rather than as a live signal that the running simulation continuously absorbs. By the stricter standard of live empirical feedback, the literature has yet to produce a single fully realized example.

The paper proposes a modular implementation with four core components: Agents (stateful decision processes with roles, objectives, beliefs, memory, tools, skills), Economic environments (runtime storing state and executing transitions, with mechanisms such as auctions, order books, matching rules, bargaining protocols, and constraints such as budgets, inventories, accounting identities, regulatory limits), Co-evolution (bidirectional improvement loop between agents and the world — "Better agents improve the world by producing harder trajectories, exposing missing rules, stress-testing mechanisms... Better worlds improve agents by providing structured feedback, counterfactual rollouts, rewards, constraint signals, memories, and stress cases"), and Real-time alignment (anchors the artificial economy to empirical, institutional, safety, or task-specific evidence during rollout).

The paper specifies a reusable implementation protocol with specification interfaces (ewm.agent, ewm.environment, ewm.coevolution, ewm.alignment, ewm.evaluation) and runtime interfaces (world.reset, world.run agents, world.step, world.coevolve, world.align, world.evaluate). A minimal usage example constructs an FX market with households, firms, and banks; each step calls world.run agents(state, parallel=True) to collect actions, then world.step(actions) to advance the economy.

Evaluation is layered across five targets: agent-level metrics (action validity, role consistency, belief calibration), environment-level metrics (constraint violation rate, market-clearing error, accounting consistency), co-evolution metrics (adaptation gain, stability, drift), real-world alignment metrics (state error, trend match, price/volume/volatility error), and efficiency metrics (runtime cost, memory usage, parallel efficiency).

The paper traces the technical trajectory toward EWMs through five engineering stages:

1. Feature engineering — designing observation signals: It converts raw economic traces into structured signals that agents can use to perceive the economic state. Examples include Gu et al. (2020) on predictive signals for asset pricing, Manela and Moreira (2017) on text-based uncertainty indices, and Obaid and Pukthuanthong (2022) on news photos as sentiment measures.

2. Data engineering — the focus shifted from hand-designed features to large-scale data collection, cleaning, alignment, and pretraining — e.g., FinBERT, BloombergGPT, FinGPT. Data engineering strengthens the model intelligence that powers agents.

3. Prompt engineering — crafting precise input instructions to elicit sophisticated economic reasoning. Techniques include few-shot examples, role-playing, and chain-of-thought prompting. Prompt engineering becomes relevant only once agents acquire a language-based cognitive substrate, marking the transition from L2 to L3.

4. Context engineering — engineers the broader context available to LLMs during reasoning via retrieval-augmented generation, structured memory, and tools. Context engineering evolves from memory and tool use at L3, to reflection at L4–L5, and finally to real-time empirical grounding at L6.

5. Environment engineering — the economy itself becomes the object of design... the wave that turns models into worlds. Examples include ABIDES (Byrd et al., 2020), AI Economist (Zheng et al., 2022), TaxAI, EconAgent, and StockSim. "At L1, it defines fixed world rules. At L2, it provides feedback channels for adaptive agents. At L3, it offers execution interfaces for LLM-based agents. At L4 and L5, it supports interaction loops through which agents and institutions can co-evolve. At L6, it provides the alignment machinery that links simulation states to real-world data."

The paper organizes applications into three functional roles:

1. EWM as a sandbox for human decision making — supporting policy simulation and stress testing, financial market and systemic risk analysis (liquidity dry-ups, deleveraging spirals, fire-sale externalities, contagion), firm strategy and supply chain resilience, and research and education.

2. EWM as a tool for LLM agents — providing counterfactual economic reasoning (an agent can query the world model to compare possible outcomes under endogenous responses from other agents), long-horizon planning grounded in accounting identities, market-clearing conditions, and institutional structure, and safety and deployment evaluation (testing for collusion, instability, manipulation, harmful equilibria, and cascading failures before deployment).

3. EWM as an RL training environment for LLMs — supplying states, actions, transition dynamics, and both short- and long-term economic reward signals embedded in environments shaped by scarcity, institutional rules, heterogeneous counterparties, and strategic interaction. Example: AlphaPortfolio (Cong et al., 2021).

The paper contrasts EWMs with four adjacent paradigms:

• Traditional economic models (microeconomic equilibrium, macroeconomics, CGE, DSGE): Traditional models close the system through equilibrium and rational expectations and rely on representative agents. EWMs preserve the commitment to consistency but implement it through explicit system components with heterogeneous agents, hybrid state representations pairing structured variables with language-based memory, and beliefs as observable intermediate objects. However, for counterfactual analysis, "capability is not enough: equilibrium discipline, and in learned environments DDGE-style closu…

Based on the paper, here are concrete improvements to AI systems and what the improved systems can do:

---

1. Close the action–consequence loop.

• Improvement: Replace stand-alone next-action prediction with an engineered closed loop between agent-side action generation and world-side next-state prediction. Implement world.run agents followed by world.step, where actions are realized, evaluated, and returned as learning signals.

• Can do: An AI agent can take actions in a simulated economy, observe the market-clearing consequences, receive constraint and reward feedback, and update its policy — instead of predicting actions without experiencing outcomes.

2. Model heterogeneous agents with explicit beliefs and private states.

3. Implement typed economic states.

4. Make aggregation endogenous.

5. Validate behavior instead of relying on persona prompts.

6. Enable co-evolution of agents and institutions.

7. Add real-time empirical alignment.

8. Support a six-level capability progression.

9. Decompose dynamics into five explicit operators.

10. Use multi-target evaluation.

11. Build counterfactual reasoning and long-horizon planning for LLM agents.

12. Use the economy as an RL training environment.

13. Maintain equilibrium discipline even with learned components.