Source-linked AI summary
But How Would AI Agents Run a Town's Economy?
Sajal Regmi, Siddhartha Pudasaini, Chetan Phakami Pun
TL;DR
Prior LLM-agent economy studies often stop after short runs and report plausible behavior rather than whether resources move through an economy. This paper follows a closed, money-conserving simulation for up to 26 weeks and finds that demand reaches firms far more than workers, while wealth mobility changes with the horizon.
Problem
Many agent-society studies run for days or a couple of simulated weeks and emphasize plausible behavior, with limited testing of whether aggregate economic outcomes react to shocks.
Source: §1 · p. 1Method
Results
A 12× tourist shock raises revenue 4.62× while wages move 1.03×; wealth persistence falls from ρ=0.964 at two weeks to 0.752 at 26 weeks.
Source: §1 · p. 1Takeaways & Limitations
Without incentives to move wages or prices when shifts sell out, windfalls concentrate at firms by construction, regardless of the underlying LLM’s capability.
Source: §10 · p. 7The horizon comparison rests on nine two-week runs, four 12-week runs, and a single 26-week path, so the longest-horizon result is not a distributional estimate.
Source: §11 · p. 8
Abstract
from arXiv · showhide
We placed 100 memory-equipped large language model (LLM) agents in charge of a closed, money-conserving spatial economy on real Pokhara Lakeside geography (earning wages, running businesses, setting prices) and ran this multi-agent simulation for up to 26 simulated weeks, well past the 1-2 weeks typical of agent-society studies. Across 91 validated runs (2.44M agent decisions, 21.5B tokens), the money stops moving, in a specific and measurable way. A 12x tourist demand shock raises business revenue 4.62x ($p<0.001$), which we decompose exactly into a 1.50x extensive margin (more businesses trading) and a 3.07x intensive margin (more revenue each). Monetary transmission stops there. Wages move 1.03x ($p=0.42$); 0.3% of 3,981 menu items are ever repriced ($p=0.47$). A randomized cash transfer (NPR 5,000 to 20 of 100 agents) shows the same pattern from the opposite direction: 96.7% is still held 311 pulses later, marginal propensity to consume 3-4% by two independent measures, indistinguishable from zero. The wealth distribution is consequently near-frozen at the horizon this literature uses ($\rho=0.964$ over 2 simulated weeks), but not frozen. $\rho$ falls to 0.832 at 12 weeks and 0.752 at 26, a horizon-dependence no short study can see. Matched ablations show which knob actually matters. Swapping the backing LLM moves every outcome we measure ($p=0.0039$); deleting agents' memory moves none of them detectably. A purely social tool fails 94-97% of the time across two model families, compared with ~96% success on economic tools, with no measurable shift away from it. Every headline number is verified twice, by a live validator and by an offline recomputation that reconciles each agent's wealth against its own signed transaction history, and we release the full run corpus for reanalysis.
1 Introduction
The paper tests whether economic shocks and windfalls circulate through an LLM-agent economy, using a validated, money-conserving simulation run for up to 26 weeks. Demand reaches firms but barely reaches workers, household transfers are largely retained, and wealth mobility depends on the observation horizon.
1 Introduction
- The study extends prior agent-economy work by tracking resource flows for up to 26 simulated weeks rather than stopping at the usual one to two weeks.The closed system conserves money exactly across 41,328 recorded pulses in all 91 validated runs.Source: §1 · p. 1
- A 12× tourist shock raises revenue 4.62×, decomposed into a 1.50× extensive margin and a 3.07× intensive margin, while wages move only 1.03×.Only 0.3% of 3,981 menu items are ever repriced.Source: §1 · p. 1
- 96.7% of a randomized transfer remains held 311 pulses later, with independent marginal propensity to consume estimates of 3.3% and 4.0%, neither distinguishable from zero.The transfer was NPR 5,000 to 20 of 100 agents.Source: §1 · p. 1
- Wealth persistence falls from ρ=0.964 at two weeks to 0.832 at 12 weeks and 0.752 at 26 weeks, with the Gini declining from 0.674 to 0.643 by week 16.Only 21% of agents ever change quintile at the conventional two-week horizon.Source: §1 · p. 1
- Swapping the backing LLM moves every measured outcome, while removing agents’ memory moves none detectably; invite_to_talk fails 94–97% of the time across two model families.The social tool accounts for 79.6% of all tool failures.Source: §1 · p. 1
- The study contributes 91 validated runs, an extensive/intensive decomposition of monetary transmission failure, a randomized transfer test, and evidence that wealth mobility depends on horizon.The run corpus is released, but the platform is not.Source: §1 · p. 1
2 Related Work
Prior agent-society studies report plausible behavior, while economic-agent and benchmark work generally examines narrower tasks or outcomes. This paper positions ledgered, long-horizon macroeconomic measurement as its contribution.
2 Related Work
- Prior town, market, and organization simulations commonly run for days or a couple of weeks and emphasize recognizably human behavior rather than resource movement.The paper asks whether demand reaches wages, windfalls are spent, and wealth evolves over longer horizons.Source: §2 · pp. 1–2
- Economic-agent studies often probe surveys, auctions, bargaining, autonomous trading, or narrow mechanisms rather than sustaining a 100-agent economy over weeks.EconAgent and EconSimulacra come closest, but do not conserve money at transaction level, which this paper uses for exact margin decomposition.Source: §2 · p. 2
- Human households’ marginal propensity to consume is typically 0.2–0.5, against which the paper’s measured 3–4% is described as a sharp outlier.The comparison is framed around consumption out of a windfall.Source: §2 · p. 2
- Unlike per-episode benchmarks such as AgentBench and WebArena, this setting follows a closed economy’s aggregate trajectory across thousands of decision rounds.The paper reports macroeconomic and distributional measures such as Gini, wealth-rank persistence, and wage pass-through instead of task-level scores.Source: §2 · p. 2
- The authors frame the work as an empirical multi-agent-systems contribution that identifies a mechanism behind a specific failure, not a capability demonstration.They present the conserved-money ledger and two independent validators as an answer to validation concerns in generative agent-based modeling.Source: §2 · p. 2
3 The Environment
The environment places LLM agents in a geographically grounded town with businesses, asynchronous decision-making, logged tools, and exact money accounting.
3 The Environment
- The Lakeside, Pokhara registry contains 762 places across seven categories, with 735 priced menus totaling 3,981 items.One hundred agents have cash wallets, bank balances, and an 80 m visibility radius.Source: §3 · p. 2
- World time advances in fixed 60-simulated-minute pulses, while agents wake, reason, and act asynchronously within each pulse.A 336-pulse two-week run is a staggered stream of reasoning episodes, not 336 simultaneous rounds.Source: §3 · p. 2
- Agents share a fixed 19-action toolset covering navigation, economic activity, memory, and social coordination.Business owners can set prices and pay wages, and tool calls, transactions, price observations, and conversations are logged with pulse-level timestamps.Source: §3 · p. 2
- Total agent wallet and bank money is conserved exactly, except for two ledgered external accounts: a tourist inflow pool and a government sink.Money leaving agents’ hands accrues to business cash-on-hand and bank balances.Source: §3 · p. 2
4 Experimental Design
The experimental design varies tourist arrivals, transfers, layout, memory, model, and run length across arms that share a pre-registered seed pool. The realized sweep had fewer seeds than planned and used matched-seed or within-run contrasts.
4 Experimental Design
- The design includes baseline, low/high tourism, a low-to-high-to-low tourism shock, and a wealth grant, plus layout, memory, cross-model, and extended-horizon arms.The grant goes to 20 of 100 agents at pulse 24; baseline arrivals are 2.0 per simulated hour.Source: §4 · p. 2
- The planned sweep was five conditions by 30 seeds, with a pre-set fallback of 15, but the realized sweep used nine seeds per condition and a 336-pulse horizon.The shift to self-hosted inference also changed the planned model and ablations; comparisons therefore use matched seeds or within-run contrasts.Source: §4 · p. 2
- Of 98 dispatched runs, 91 were retained; six cross-model runs halted within four pulses, and one completed tourism-low run was excluded for falling below the pre-registered arrival floor.The excluded run had 56 tourist arrivals against the 56.5 three-sigma floor.Source: §4 · p. 2
- Figure 1a depicts a transient 14× rise in tourists present, while Figure 1b summarizes 91 accepted runs across seven arms sharing a seed pool.The shock treatment’s panel reports the mean of nine seeds.Source: §4 · p. 2, Figure 1 · p. 3
- Figure 2 renders a run on real Lakeside, Pokhara geography, with agents positioned among 762 registered businesses.The markers show agents at their current place.Source: Figure 2 · p. 3
5 Validation
The paper validates runs with live and offline checks, including per-agent ledger reconciliation, and shows that money balances exactly across recorded pulses. Table 1 reports that 17 of 19 checks pass on all 91 runs, while serving and probe conditions qualify tool-reliability results.
Validation pipeline
- Live TypeScript checks and an independent offline Python validator recompute headline numbers from raw exports, including per-agent ledger reconciliation.The reconciliation requires each agent’s wallet-plus-bank balance change to equal the signed sum of its transaction history.Source: §5 · p. 3
Validation results
- 17 of 19 checks, including ledger reconciliation and every transaction-flow total, pass on all 91 runs with zero exceptions.Table 1 also flags a benign Gini snapshot staleness issue and a token-counter mismatch on gpt-oss runs.Source: §5 · p. 3, Table 1 · p. 3
Serving differences
- 73 of 91 token-ledger checks fail on gpt-oss runs and none on Qwen runs because malformed generations enter the pulse counter but not the per-call ledger.After serving fixes, 0.86% of gpt-oss calls remain malformed, versus 0.00% for Qwen.Source: §5 · p. 3
Money conservation
- Total money equals the sum of agent balances at all 41,328 recorded pulses across 91 runs, with zero deviation.The conserved-money identity supports the margin decomposition and transfer tracing.Source: §5 · p. 3
Tool-reliability probes
- A reduced tool-reliability probe reported 100% success on the failure mode and would have concealed the §9 result.The passage recommends replaying recorded production requests for reliability probes.Source: §5 · pp. 3–4
6 Transmission Failure
A controlled tourism-intensity sweep shows that demand growth raises business revenue but does not pass through to wages or prices. The revenue response decomposes into more businesses trading and higher revenue per active business, without a mechanical link to worker pay or menu prices.
6 Transmission Failure
- Tourist spending separates completely across the intensity sweep (δ=1.00, p_adj<0.001), so downstream null results are not attributed to a weak treatment..Source: §6 · p. 4
- Business revenue rises 4.62× across the tourism sweep, while wages remain flat and only 0.27–0.32% of 3,981 menu items are repriced.Wages are 260,389 versus 267,459 NPR (p=0.42); repricing is unresponsive to a 4.6× revenue swing (p=0.47).Source: §6 · p. 4
- The 4.62× revenue response decomposes into 1.50× more businesses trading and 3.07× more revenue per active business.Those margins multiply to 4.61×, matching the observed 4.62× up to rounding; 458 versus 304 of 762 businesses trade at high demand.Source: §6 · p. 4
- Agents rarely call set_price—about 18 times per run against roughly 28,000 total calls—and initial menus are sticky by construction.Under the policies agents adopt, demand gains concentrate at owners rather than passing to workers or customers.Source: §6 · p. 4
7 Windfalls Are Not Circulated
A within-run randomized cash transfer tests whether money reaching households directly circulates. Most of the grant remains held, and two spending-based estimates of consumption response are indistinguishable from zero.
7 Windfalls Are Not Circulated
- Twenty of 100 agents were randomly assigned NPR 5,000 at pulse 24, with the remaining 80 serving as controls in the same world.The study tracks the wealth gap across 336 pulses and measures post-transfer purchase spending.Source: §7 · p. 4
- The authors interpret the agents as aggregate savers; a windfall that does not circulate cannot generate second-round demand, part of why wealth moves slowly..Source: §7 · p. 4
8 The Wealth Distribution Only Looks Frozen
Wealth persistence is high at the conventional two-week horizon but falls as the simulation runs longer. The long-horizon evidence is based on five runs, with the 26-week result coming from a single trajectory.
8 The Wealth Distribution Only Looks Frozen
- At the conventional 2-week horizon, wealth persistence is ρ=0.964 and only 21% of agents ever change wealth quintile.Persistence remains ρ=0.964–0.970 across the tourism sweep.Source: §8 · p. 4
- Wealth persistence falls from ρ=0.964 at 2 weeks to 0.832 at 12 weeks and 0.752 at 26 weeks, while quintile mobility rises from 21% to 45% and 46%.The single 26-week trajectory shows the Gini falling sharply by week 10 and plateauing near 0.643.Source: §8 · p. 4
- The 12- and 26-week results rest on five runs, and the 26-week trajectory is a single path.The authors say wealth-distribution conclusions are only as trustworthy as the measured horizon.Source: §8 · p. 4
9 What Actually Drives Behavior
Matched-seed contrasts test which controlled factors shape behavior; model identity matters, while memory and layout show no detectable economic effect.
Placebo: world layout
- Changing only the world layout leaves Gini, persistence, and other outcomes statistically indistinguishable from baseline.Gini changes +0.1% (p=1.00) and persistence +0.2% (p=0.84).Source: §9 · p. 5
Model identity and memory
- Swapping the LLM shifts outcomes 3–10× more than withdrawing memory, while model identity matters and memory has no detectable effect among measured outcomes.The figure compares matched-seed contrasts against baseline core-a.Source: §9 · p. 6, Figure 6 · p. 7
Social coordination tools
- Social tools fail 94–100% of the time under Qwen3.8-27B, while economic and navigation tools mostly succeed.The platform concurrency cap is the dominant logged failure reason, not evidence about LLM social cognition.Source: §9 · p. 6, Figure 7 · p. 7
10 Discussion
The results converge on a system where agents earn and hold money without spending it, while the wealth distribution relaxes only over longer horizons. For economies intended to resemble human ones, the discussion prioritizes transmission incentives over memory or governance.
10 Discussion
- Across four independently identified results, demand reaches firms but not households, transfers do not circulate, wealth mobility emerges with time, and outcomes vary with the backing LLM rather than memory.Together, these results describe agents that earn and hold money without spending it at every tested scale.Source: §10 · p. 7
- Without incentives to adjust wages or prices when shifts sell out, the system concentrates windfalls at firms regardless of the underlying LLM’s capability.The discussion identifies transmission failure as the first issue to address for an economy intended to behave like a plausible human one.Source: §10 · p. 7
- Distributional claims from LLM-agent societies should report their horizon, because short-horizon stability may reflect stopping before wealth relaxes.An aggregate success rate can also conceal near-total, model-agnostic failure of a particular action class.Source: §10 · p. 8
- The authors propose falsifiable tests: a prompting change that restores pass-through would challenge the transmission claim, while longer-run replicates could test the horizon ordering.The prompting intervention was not tested.Source: §10 · p. 8
11 Limitations •
The evidence is constrained by small long-horizon samples, an underpowered memory comparison, price-setting behavior partly shaped by design, and a single-town, single-stack setup.
11 Limitations
- Long-horizon evidence rests on 9 runs at 2 weeks, 4 at 12 weeks, and one 26-week path, not a distributional estimate.The authors rely on the ordering that every 12-week run falls below every 2-week run.Source: §11 · p. 8
- The memory comparison has only 6 matched seeds, so no detected effect does not establish equivalence.Source: §11 · p. 8
- Near-zero repricing reflects both rare use of the available set_price lever and a menu designed to be sticky; these influences were not separated.set_price is called about 18 times per run.Source: §11 · p. 8
- The study varies 9 agent-draw seeds on one map, uses two models on self-hosted vLLM, and includes no governance.Mistral-Small-3.2-24B could not complete a single pulse, which the authors describe as stack-specific rather than a general model result.Source: §11 · p. 8