100 AI Agents Ran a Town for 26 Weeks. The Money Stopped Moving.

A 12× demand shock raised business revenue 4.62× and wages 1.03×. What a released LLM economy tells us about multipliers, horizons and which knob actually drives agent behavior.

A team at Karela Technologies put 100 memory-equipped LLM agents in charge of a tourist town modeled on Lakeside, Pokhara, and ran it for up to 26 simulated weeks. The agents worked shifts, ran 762 businesses and set prices inside a closed economy where every rupee is ledgered. Then they released the full run corpus: 98 runs, about 7.8 GB, under CC BY 4.0.

Below are the four ideas that matter most from the paper, with the numbers. The full deep dive on Substack has the derivations, a relaxation model I fit to the horizon data, and Python to recompute the distribution metrics from the released panel yourself.

A Multiplier of Almost Exactly One

The cleanest frame for every result is the Keynesian spending multiplier, k = 1 / (1 - c), where c is the marginal propensity to consume. The paper gave NPR 5,000 each to 20 randomly chosen agents and tracked the money. About 13 simulated days later, 96.7% was still unspent. Two independent estimates put the agents' MPC at 3.3% and 4.0%. Human households typically run between 0.2 and 0.5.

Plug in c = 0.04 and k ≈ 1.04. The NPR 100,000 grant generates about NPR 4,167 of spending for anyone other than the recipients. At c = 0.5, the same grant generates NPR 100,000. A human economy passes 6 to 24 times as much money downstream.

The full article walks through why the multiplier, crude as it is in real macro work, fits a closed ledgered economy unusually well.

Demand Stops at the Firm

A 12× tourist shock lifted business revenue 4.62×. Because money is conserved, the authors split that exactly: more firms trading (304 to 458, about 1.51×) times more revenue per firm (3.07×). Then the chain stops. Wages moved 1.03×, and the wage share of revenue fell from 0.90 to 0.20.

The implied wage elasticity to revenue is ln(1.027) / ln(4.620) ≈ 0.0175. Double revenue and wages rise about 1.2%. Prices barely moved either: only 0.27% to 0.32% of menu items were ever repriced, and set_price accounts for about 0.06% of tool calls. The lever exists. Agents almost never pull it.

The full article covers what this means for anyone building agent systems: an action available in a toolset is not an action the policy will take.

The Distribution Only Looks Frozen

At the two-week horizon most agent studies use, wealth ranks barely move: Spearman persistence is 0.964 and only 21% of agents change quintile. At 12 weeks persistence falls to 0.832, and at 26 weeks to 0.752, with 46% changing quintile.

I fit a simple decay model, rho(t) = rho_inf + (1 - rho_inf) * exp(-t / tau), to those points. It gives a half-life of about 10.2 weeks. That fit rests on very few runs, including a single 26-week path, so treat it as a design heuristic: to see most of a distributional adjustment in a system like this, plan for 20 to 30 simulated weeks, not two.

The full article includes the fit code and a recipe for refitting it week by week on the released long run.

Model, Not Memory

The ablations are the most practical result. Removing memory entirely moved no economic outcome detectably. Swapping the decision model from Qwen3.8–27B to gpt-oss-20b moved every outcome at p = 0.0039, including median wealth −22.9% and quintile mobility +62.4%.

So the agents are not reciting the textbook (which would raise wages and spend windfalls), and they are not clearly reinventing economics from interaction. Outcomes track the policy and the action interface. A social tool that failed 97.2% of the time under Qwen kept getting called anyway, with no measurable adaptation.

The full article turns this into reporting rules for multi-agent work: treat the model and serving stack like a seed, and log failure per tool.

What’s in the Full Article

• The multiplier frame, with every worked number recomputed in Python
• The exact revenue decomposition and the wage-elasticity calculation
• A proposed relaxation model with a 10.2-week half-life, plus fit code
• Runnable Python for Gini, top-10% share, rank persistence and quintile mobility on the released panel
• A frank “where it breaks” section: 9 seeds instead of 30, one 26-week run, and the prompt test the authors did not run

For ML engineers, economists and anyone designing agent simulations who wants numbers they can check.

Read the full deep dive on Substack → https://brianjamescurry.substack.com/p/do-llm-agents-reinvent-economics

Subscribe on Substack for weekly deep dives with code and math you can run.

About the Author

Brian Curry is a Kansas City-based machine learning and AI strategist, marketing economist, and applied researcher. He leads Vector1, an independent research studio working on cognitive AI, causal inference, marketing economics, and knowledge engineering, with a focus on simulation-driven intelligence systems.

Over more than twenty years, Brian has led data science and analytics initiatives at enterprises ranging from $4B to $125B in revenue, including Koch Industries, Tractor Supply, Vail Resorts, Hallmark, Garmin, and AT&T. His work applies economics and causal inference to connect marketing investment to structural outcomes, operational decisions to economic value, and AI adoption to durable competitive advantage.

He is also the founder of KC AI Lab, a 1,500-member AI and ML education community in the Midwest. He lives in Kansas City and is an active part of the Kansas City community.

Research and Open-Source Projects

The modeling approach in this paper draws on PyCausalSim and Papilon.

Causal Inference and Economics

• PyCausalSim: Simulation-driven causal discovery and evaluation, using counterfactual scenarios to isolate structural drivers.
• Papilon: Complex-system simulation and optimization, with applications in media mix modeling and adaptive architectures.
• Daedalus: Causal analytics for money-flow mapping and macroeconomic tracking.
• The Autonomy Economy: A macroeconomic framework for the combined effects of rapid productivity growth and structural price deflation in an AI-automated economy.

Forecasting and Applied Modeling

• Regimecast: Time-series regime detection, structural breaks, and economic forecasting.
• CarryNet: Financial modeling for commodity carry trade analysis.
• Foodmesh-ml: Machine learning, optimization, and simulation for distributed food systems.

Cognitive AI and Semantic Architecture

• Memory-Node Encapsulation (MNE): A cognitive AI architecture and data structure for artificial episodic memory (NeoCortex-M).
• MeaningFlow: Semantic content modeling, embeddings, clustering, and coverage gap analysis.

Contact: brian@vector1.ai | GitHub @Bodhi8 | Substack

The text of this article may be shared with attribution under Creative Commons Attribution 4.0 International (CC BY 4.0). Code and models referenced here are available separately to paid subscribers.