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

Economic Simulations That Learn and Adapt

Economic world models—generative simulations that evolve from the ground up by modeling individual agents, their beliefs, decisions, and interactions—could become powerful tools for understanding and stress-testing economies. The authors argue that building truly useful economic simulations requires a systematic roadmap, and they lay out a six-level ladder of increasing capability that reveals a critical gap: most existing work stops at simple, rule-based agent worlds, while the ambitious goal of self-adapting agents, endogenous institutions, and real-world calibration remains largely unexplored.

How It Works: A Ladder of Complexity

The authors organize economic world models into six ascending levels:

1. Fixed rule-based agent worlds: Agents follow predetermined behavior rules in simple market environments.
2. Adaptive agent worlds: Agents learn and adjust behavior over time without large language models.
3. LLM-based agent worlds: Language models power agent reasoning and decision-making.
4. Self-evolving agents: Agents not only adapt but generate their own strategies and beliefs endogenously.
5. Evolving institutional worlds: Markets, institutions, and rules themselves transform as agents interact—not frozen exogenously.
6. Sim-to-real economic twins: Simulations persistently aligned with real economic data and validated against actual market behavior.

The key insight is that each level compounds—moving beyond isolated agents to systems where beliefs, markets, and institutions co-evolve dynamically. A literature survey finds that published work clusters heavily in levels 1–2 (rule-based and mildly adaptive simulations), with sparse progress on levels 4–6, where agents drive institutional change or models persistently track real economies.

Why It Matters

The stakes are high. Economic simulations have traditionally been either overly abstract (general equilibrium theory) or narrowly calibrated to specific datasets with limited generalization. Economic world models promise a middle path: rich enough to capture heterogeneous agents' beliefs and strategic behavior, yet general enough to explore "what if" scenarios across policy and institutional designs.

The authors frame EWMs as serving two purposes. For human decision-makers, they become high-fidelity sandboxes to test fiscal policy, market designs, or institutional reforms without real-world risk. For AI agents, they provide training grounds and evaluation benchmarks—environments to learn economic reasoning, plan actions, or assess safety implications of autonomous economic actors.

The six-level ladder itself is the paper's main contribution: it codifies what "done well" looks like and exposes where the field is underdeveloped. Notably, the absence of systems combining self-evolving agents, endogenous institutions, and persistent empirical alignment suggests the frontier lies in building EWMs where agents don't just optimize within fixed rules—they reshape the rules through their collective behavior, and those reshaped systems remain tethered to real economic patterns.

The authors release a curated paper list and resource collection to unblock future work. This blueprint is less a specific algorithm or empirical result and more a research agenda: a structural argument that economic simulation capability has levels, that we're early in the stack, and that closing the gap between simulation and reality while permitting genuine institutional endogeneity will require deliberate, interdisciplinary effort bridging economics, AI, and complexity science.