PaperlayerView exact revision on arXiv

Source-linked AI summary

Financial Fragility in Societies of LLM Agents: Coordination Failures and Stabilizing Mechanisms

Zhenhao Fu, Ruipeng Xu, Qibing Ren

arXiv:2609.30940v1cs.AIq-fin.GN

TL;DR

Interdependent financial decisions can produce collective failure even when agents act protectively, raising a system-level safety question as LLMs take on financial roles. FRAIL tests this in three dynamic environments and compares commitment mechanisms. Failures occur across environments, while mechanisms improve outcomes and stabilization shares an early-commitment pattern.

  • Problem

    Can multiple LLM agents in the same financial system generate systemic fragility, how does it emerge, and how can it be mitigated?

    Source: §1 · p. 1

  • Method

    FRAIL places LLM agents in three dynamic financial environments and compares interaction mechanisms that alter commitment, information, and incentives.

    Source: §1 · p. 2, §5 · p. 9

  • Results

    Across three environments, individually protective decisions can generate collective failure; 77% of baseline bank-run episodes fail, and all three mechanisms improve outcomes.

    Source: §1 · p. 2, §5 · p. 9

  • Takeaways & Limitations

    Financial-agent evaluation should measure the dynamics and stability of systems agents collectively create, and interaction mechanisms should fit the financial structure.

    Source: §5 · p. 9

    Results concern simulated agents in controlled, stylized financial environments and are not forecasts of real-world financial crises.

    Source: §3.1.3 · p. 5, ETHICS STATEMENT · p. 10

Abstract

from arXiv · show

Individually protective decisions can produce avoidable collective failures. As large language model (LLM) agents take on greater roles in financial decision-making, financial AI safety must therefore be considered not only at the level of individual agents, but also at the level of the systems they jointly create. We study this problem with FRAIL, a controlled experimental framework that places LLM agents in three dynamic financial environments---bank runs, debt rollover, and reward crowdfunding---where agents' decisions reshape the financial conditions faced by others. Across seven leading LLMs, we find widespread collective fragility even when no agent is instructed to destabilize the system: 77\% of baseline bank-run episodes and 83\% of debt-rollover episodes end in failure. We then compare three interaction mechanisms based on compensated commitments, centralized commitment agreements, and participant-led coalitions. All three improve aggregate outcomes, but no single mechanism performs best across all financial structures. Across mechanisms, successful stabilization shares a common temporal pattern: broad commitment forms early, before defensive behavior becomes self-reinforcing. Our findings show that individually capable agents do not automatically form safe financial systems, highlighting system-level evaluation and interaction design as central problems for financial AI safety. Code is available at https://anonymous.4open.science/r/FinFrail-CF26.

1 INTRODUCTION

Financial decisions can be individually protective yet collectively destabilizing when actions reshape conditions for others. FRAIL tests whether this fragility emerges across LLM agents and compares commitment mechanisms for mitigation.

1 INTRODUCTION

  • The paper asks whether LLM agents sharing a financial system can generate systemic fragility, how it emerges, and how it can be mitigated.The motivation is that one participant’s actions can change the prices, liquidity, or opportunities faced by others.Source: §1 · pp. 1–2
  • FRAIL evaluates LLM agents in bank runs, debt rollover, and reward crowdfunding, where actions dynamically reshape shared financial conditions.A simulation engine maintains balance sheets, enforces commitments, and controls what participants observe.Source: §1 · p. 2
  • 77% of baseline bank-run episodes and 83% of debt-rollover episodes fail; 23% of crowdfunding episodes fail to launch, with failures spanning model families.These outcomes occur without instructions to destabilize the system.Source: §1 · p. 2
  • 26% is the average bank-run failure rate under participant-led coalitions, down from 77%; stabilization depends on broad commitment forming before defensive behavior self-reinforces.All three commitment mechanisms improve outcomes, and their effectiveness is robust to parameter ablation.Source: §1 · pp. 2–3

2 RELATED WORK

Prior work examines cooperation and interaction risks among LLM agents, including institutional mechanisms for sustaining cooperation. FRAIL focuses on settings where agents’ actions reshape financial states and incentives.

2 RELATED WORK

  • Cooperation studies report variation across models, partners, prompts, and interaction structures; related mechanism comparisons find contracting and mediation particularly effective among capable models (Piatti et al., 2024; Tewolde et al., 2026).The broader literature considers repeated games, economic games, common-resource dilemmas, communication, reputation, mediation, and contracting.Source: §2 · p. 3
  • Multi-agent safety research identifies miscoordination, conflict, and collusion, while FRAIL examines how institutional mechanisms reshape the dynamics producing interaction failures (Hammond et al., 2025; Cemri et al., 2025).The mechanisms vary in who initiates commitments, whether transfers are used, how commitments bind, and what information becomes visible.Source: §2 · p. 3
  • Finance-focused LLM studies examine social contagion, withdrawal cascades, and collusion; FRAIL compares institutional interventions across three financial coordination structures (Rua˜no & Rajan, 2026; Ren et al., 2026).Unlike fixed-payoff settings, actions in FRAIL reshape financial states and the incentives facing later decisions.Source: §2 · p. 3

3 DESIGN OF FRAIL

FRAIL uses three calibrated financial environments to study how interdependent decisions affect outcomes, then compares commitment mechanisms and evaluates agents across models and episodes.

3.1.1 BANK RUN

  • In bank runs, fifteen depositors make withdrawal decisions over twelve rounds; discounted asset sales fund withdrawals, and liquidity failure or insolvency ends in bank failure.Withdrawals can prompt further withdrawals by weakening the bank through costly liquidation.Source: §3.1.1 · p. 4

3.1.2 DEBT ROLLOVER

  • In debt rollover, eight creditors decide across four maturity waves whether to demand repayment or roll over; early repayment can require discounted asset sales and lead to default.The seven-round environment models claims maturing in rounds 1, 3, 5, and 7.Source: §3.1.2 · pp. 4–5

3.1.3 REWARD CROWDFUNDING

  • In reward crowdfunding, fifteen investors can pledge, wait, or withdraw; the founder can launch at fourteen units, while unlaunched pledges are refunded at the deadline.Investors may wait because their capital earns no outside return while the pool is underfunded.Source: §3.1.3 · p. 5
  • Across the three environments, parameters make coordinated outcomes viable while leaving failure possible under decentralized play, a region termed the coordination threshold.This calibration makes collective outcomes sensitive to agents’ decisions.Source: §3.1.3 · p. 5

3.2 INTERACTION MECHANISMS

  • Against a no-mechanism baseline, M1 uses compensated commitments, M2 uses centrally solicited unpaid agreements, and M3 uses participant-led coalitions.Participation is voluntary; while a commitment is active, corresponding funds cannot exit, and disclosures are certified.Source: §3.2 · p. 5

3.3 EVALUATION SETUP

  • Seven models spanning six families are evaluated with five episodes per model–environment–mechanism cell; funding providers report action probabilities that the environment samples.Outcomes include episode success and normalized social welfare, scaled between environment-specific worst- and best-case totals.Source: §3.3 · p. 6

4 RESULTS

Across bank runs, debt rollover, and crowdfunding, LLM agents show collective fragility, while commitment mechanisms improve outcomes without a universally best design. Stabilization is associated with early, broad commitment, and the main mechanism patterns persist across shifts in economic parameters.

4.1 BASELINE FAILURE AND MECHANISM EFFECTIVENESS

  • Only 23% of baseline bank-run episodes and 17% of debt-rollover episodes succeed, while 77% of crowdfunding episodes launch.Four of seven models lose every baseline bank-run episode; defensive exits deplete shared liquidity in bank runs and debt rollover.Source: §4.1 · p. 6
  • All three mechanisms improve welfare and success in every environment, but M3 leads in bank runs and crowdfunding while M2 leads in debt rollover.Figure 3 summarizes welfare realization and mechanism performance; the ranking also varies across models with actual mechanism utilization.Source: §4.1 · pp. 6–7, Figure 3 · p. 7
  • M1 raises participant payoff from 0.398 to 0.688, but central-player payoff stays near baseline at 0.347 versus 0.345; transfer-free M2 and M3 raise it to 0.624 and 0.668.M2 and M3 also produce broader early participation than M1, at 0.570 and 0.601 versus 0.299.Source: §4.1 · p. 7

4.2 HOW DO MECHANISMS ACTUALLY HELP?

  • Stabilization succeeds when commitment arrives early enough to redirect financial trajectories before defensive behavior becomes self-reinforcing.Many debt-rollover failures occur before the borrower’s first assets mature, and baseline crowdfunding reaches only 6.3 of 14 required units after round 1.Source: §4.2 · p. 8, Figure 4 · p. 9
  • Episodes with committed share above 45% in the first three rounds succeed 82% of the time, versus 37% when commitment stays at or below 10%.Across environments, early commitment is associated with greater success or earlier launch; uncommitted bank-run agents’ withdrawal probability stays near 20% under M3 while M0 rises toward 53%.Source: §4.2 · p. 8
  • M2 and M3 build larger commitments without transfers: bank-run committed amounts average 10.2 under M2 and 11.9 under M3, versus 8.9 under M1.The passage links paid offers to upfront balance-sheet costs and smaller commitments, while broader early commitments under M2 and M3 accompany stronger participant confidence.Source: §4.2 · p. 8

4.3 ROBUSTNESS AND ABLATION ACROSS ECONOMIC REGIMES

  • Across economic regimes, M3 leads bank runs and M2 leads debt rollover; crowdfunding is the exception, with M2 slightly ahead under expansion.Under expansion, only 3 of 20 bank-run episodes succeed; under contraction, M3 succeeds in 15 of 20 bank-run episodes and M2 in 16 of 20 debt-rollover episodes.Source: §4.3 · p. 8, Figure 4 · p. 9

5 DISCUSSION AND CONCLUSION

FRAIL finds that protective decisions by LLM agents can still produce collective financial failure, and that stabilization depends on early broad commitment suited to the financial structure. The discussion calls for system-level evaluation alongside model evaluation; the accompanying statements describe AI use, ethics, reproducibility, and glossary terms.

5 DISCUSSION AND CONCLUSION

  • Across three financial environments, individually protective decisions can generate collective failure, while structured interaction mechanisms substantially improve outcomes.Successful stabilization depends on broad commitment forming before fragility becomes self-reinforcing.Source: §5 · p. 9
  • Baseline failures across models and differing mechanism benefits show why financial AI evaluation must measure system dynamics and stability, not only local decisions.Each agent reshapes the state faced by later decision-makers; individual capability does not guarantee system stability.Source: §5 · p. 9
  • Stabilization shares an early-commitment pattern, but no mechanism is best across all environments because their timing, funding direction, and liability structures differ.The mechanism that best fills vulnerable early rounds differs by financial structure.Source: §5 · p. 9

AI USE STATEMENT

  • The authors report using generative AI for writing, retrieval, research ideation and execution, drafting, and assistance with mathematical proofs, and say they reviewed the work.They take responsibility for the final content and AI-assisted artifacts.Source: AI USE STATEMENT · p. 10

ETHICS STATEMENT

  • The study uses simulated agents in stylized financial environments, with no human subjects, personal or proprietary data, or real financial transactions; its results are not crisis forecasts.Source: ETHICS STATEMENT · p. 10

REPRODUCIBILITY STATEMENT

  • The authors release simulation code, configurations, and prompts, and report that an independent analysis pipeline can recompute all paper statistics from raw configurations and event logs.Source: REPRODUCIBILITY STATEMENT · p. 10

A GLOSSARY OF FINANCIAL TERMS

  • The glossary supplies plain-language definitions of financial terms and their counterparts in the paper’s environments.It covers concepts including coordination failure, liquidity, insolvency, maturity mismatch, and default.Source: A GLOSSARY OF FINANCIAL TERMS · p. 12

Reward crowdfunding.

  • In the crowdfunding environment, pledges are refunded if the funding goal is missed by the deadline; the glossary also defines welfare realization and the coordination threshold.Welfare realization scales social welfare between the environment’s worst and best totals; the threshold describes a region where coordinated play succeeds more comfortably than uncoordinated play.Source: Reward crowdfunding · pp. 12–13

B THE SIMULATION ENGINE

FRAIL separates LLM decision-making from financial execution: the engine controls state, validates decisions, applies mechanisms, settles outcomes, and records episodes. Its environments share a dynamic-game structure with mechanism states that shape feasible actions, information, transitions, and payoffs.

B THE SIMULATION ENGINE

  • The engine maintains authoritative financial state, gives agents authorized views, validates their structured reports, and executes rounds to produce updated states.The environments share this architecture while remaining separate packages.Source: B THE SIMULATION ENGINE · p. 13
  • Mechanisms run as engine-side state machines that track commitments and obligations, block conflicting actions, and apply exit notice and disclosure rules.Compensation enters liabilities and is included in failure tests.Source: B THE SIMULATION ENGINE · p. 13
  • The engine samples realized actions from agent-reported probabilities, applies constraints, then settles outcomes using programmatic financial rules.Settlement covers withdrawals and failure tests, maturing debt claims, or crowdfunding launches and refunds.Source: B THE SIMULATION ENGINE · p. 14
  • After execution, the engine updates certified disclosures and participant information, while logs, round checkpoints, and state hashes support replay and independent accounting checks.Replay feeds recorded decisions through the engine rather than reproducing stochastic model sampling.Source: B THE SIMULATION ENGINE · p. 14

C SHARED STRUCTURE OF FINANCIAL ENVIRONMENTS

  • The environments share a finite-horizon dynamic-game structure: funding providers and a central player act on a payoff-relevant state, with settlement, transitions, and payoffs governed by game rules.The environment is Markovian with respect to settlement, although agents’ information and policies may retain history.Source: C SHARED STRUCTURE OF FINANCIAL… · p. 14
  • A mechanism profile records commitments, exits, and transfers, and can restrict feasible financial actions; mechanisms differ in initiator, compensation, binding strength, and information visibility.Mechanism actions are resolved before same-round financial actions, and promised compensation can affect settlement liabilities and payoffs.Source: C SHARED STRUCTURE OF FINANCIAL… · pp. 14–15

D GAME PARAMETERS AND PAYOFF DEFINITIONS

This appendix specifies the environments’ parameterization, including the stable-regime settings used in the main experiments. Table 3 outlines each environment’s participants, horizon, core actions, and success or failure conditions.

D GAME PARAMETERS AND PAYOFF DEFINITIONS

D.1 PARAMETERS

D.2 REGIME VARIANTS

The appendix defines expansion and contraction variants that relax or tighten each environment’s binding constraint. It also specifies closed-form welfare reference points for the worst- and best-case outcomes in each environment.

D.2 REGIME VARIANTS

  • Expansion variants relax each environment’s binding constraint, while contraction variants tighten it; agents receive no regime label or added narrative framing.The manipulations are organized by environment and economically interpretable parameters.Source: §D.2 · p. 16, Table 5 · p. 16

D.3 WELFARE REFERENCE POINTS

  • Worst- and best-case total payoffs are derived in closed form from each configuration’s parameters.Normalized totals marginally above one are truncated to one; role-level normalized payoffs are not.Source: §D.3 · pp. 16–17
  • For bank runs, the worst-case payoff is 13.939 after a full-scale run, versus 18.175 if all depositors hold to maturity.The latter case destroys no value.Source: §D.3 · p. 16
  • For debt rollover, the worst-case payoff is 6.851; the best-case payoff is 9.899 when every claim is met and borrower profit is maximized over repayment schedules.Source: §D.3 · p. 17
  • For reward crowdfunding, the worst-case payoff is 15.450 when pledges are refunded, versus 22.802 when all investors pledge in round 1 and the founder launches immediately.Source: §D.3 · p. 17

E MODEL CONFIGURATIONS

The appendix lists the seven model configurations used in the main experiments. It also documents reasoning-effort settings and token-budget constraints that affected some model calls.

E MODEL CONFIGURATIONS

  • The main experiments use seven model configurations, listed in Table 6.The models were accessed in September 2026.Source: E MODEL CONFIGURATIONS · p. 17, Table 6 · p. 17
  • High reasoning effort repeatedly failed to terminate within the token budget, so the main experiments use low effort; MiniMax-M3 has no effort adjustment.Source: E MODEL CONFIGURATIONS · p. 17
  • MiniMax-M3 debt-rollover calls use a 16384-token limit because an 8192-token cap truncated reasoning before completion and yielded no usable result.Source: E MODEL CONFIGURATIONS · p. 17

F EXPERIMENTAL PROTOCOL

The protocol applies a common round-by-round interaction flow across environments and conditions, with mechanism actions and information delivery preceding financial decisions and settlement. The appendices document call structures, prompt contracts, mechanism-specific rules, and measurement conventions.

F EXPERIMENTAL PROTOCOL

  • Each round proceeds through mechanism actions, message delivery, financial decisions, settlement, and disclosure of the updated aggregate state.The same interaction flow applies across environments and conditions.Source: F EXPERIMENTAL PROTOCOL · p. 17
  • Agents in debt rollover and crowdfunding see the round and remaining horizon, while bank-run agents are not told either; mechanisms can deliver proposals or authorized disclosures before same-round financial decisions.Ordinary financial information reflects the prior round’s end state, so mechanism conditions combine more timely information with contractual content.Source: F EXPERIMENTAL PROTOCOL · pp. 17–18
  • Funding providers report probabilities over financial actions, which the environment samples; this permits mixed-strategy interpretation and smooth aggregation across simultaneous reporters.Source: F EXPERIMENTAL PROTOCOL · p. 18
  • The appendix identifies direct-action and equal-timeliness controls as design alternatives, alongside controls fixing role structure across mechanism conditions.The latter controls would unbundle design components combined in the present mechanism conditions.Source: F EXPERIMENTAL PROTOCOL · p. 18

G PROMPTS AND REPLY CONTRACTS

  • Prompts pair role, game-parameter, and reply-contract instructions with the previous-round state, agent position, and authorized mechanism messages; malformed replies are recorded as corrections.The appendix includes the round call structure and baseline and mechanism-specific prompt examples.Source: G PROMPTS AND REPLY CONTRACTS · p. 18, §G.1 · p. 18, Table 7 · p. 18

G.3 MECHANISM-SPECIFIC BLOCKS

  • Mechanism-specific calls encode binding commitments and group rules, including locked participation, exit rituals, and member votes on public announcements.Crowdfunding coalition membership instead locks pledges until launch, with a three-confirmation exit requirement and no bonus.Source: §G.3 · p. 21

H MEASUREMENT DETAILS

  • The measurement appendix covers episode-level uncertainty, action-level measures, scorecard dimensions, and trajectory conventions.Source: H MEASUREMENT DETAILS · p. 22

H.1 EPISODE-LEVEL UNCERTAINTY

This appendix specifies how episode uncertainty, action-level measures, scorecard dimensions, and trajectories are reported, including critical boundaries and aggregation conventions.

H.1 EPISODE-LEVEL UNCERTAINTY

  • With five episodes per cell, success rates have 0.2 resolution and wide intervals, so per-cell rankings are descriptive.Table 8 reports success rates with 95% Wilson score intervals for seven main model configurations in the stable regime.Source: §H.1 · p. 22, Table 8 · p. 22

H.2 ACTION-LEVEL MEASURES

  • Table 9 uses realized sampled actions; withdrawal and redemption rates are funding shares, while launch round averages only episodes that launch.Cells report the unweighted mean across the seven model means, with standard errors across model means.Source: §H.2 · p. 22, Table 9 · p. 23

H.3 SCORECARD DIMENSIONS

  • Figure 3b's six dimensions cover welfare, success, central-player payoff, cross-model consistency, weakest-environment welfare, and early participation.All dimensions are oriented so that larger values indicate better performance.Source: §H.3 · p. 23

H.4 TRAJECTORY MEASUREMENT CONVENTIONS

  • Trajectories mark critical boundaries; after termination, last observations are carried forward, commitments average observed episodes, and defensive intent uses only uncommitted participants.The top and middle rows show 95% confidence intervals across episodes; crowdfunding's bottom-right panel instead tracks episodes not yet launched.Source: §H.4 · p. 23

I.1 EARLY COMMITTED SHARE AND OUTCOMES

Across bank runs and debt rollover, greater early commitment is associated with higher success; in crowdfunding, it is associated with earlier launches.

I.1 EARLY COMMITTED SHARE AND OUTCOMES

  • Bank-run success rises from 0.577 in the lowest early-commitment bin to 0.906 in the highest among M1–M3 episodes.The positive association remains when baseline episodes are excluded; the bins use the maximum committed share in rounds 1–3.Source: §I.1 · p. 24, Figure 6 · p. 24
  • Across all conditions, bank-run and debt-rollover success rates rise with maximum committed share in rounds 1–3, while crowdfunding launch round falls.Figure 6 pools seven models and four mechanism conditions into five equal-width bins; launch round 10 marks episodes that never launch.Source: §I.1 · p. 24, Figure 6 · p. 24

I.2 REGIME RESULTS BY MODEL

This section reports welfare realization across economic regimes and gives four-model-subset results, including model-level results for expansion and contraction.

I.2 REGIME RESULTS BY MODEL

  • Table 10 reports welfare realization across the three economic regimes for the four-model subset.The section identifies GPT-terra, Claude-Sonnet, GLM, and Qwen 3.8 as the subset.Source: §I.2 · p. 24, Table 10 · p. 25
  • Table 11 reports welfare realization under M0–M3 for the four-model subset in the expansion regime.Arrows indicate increases or decreases relative to M0 for the same model and regime.Source: §I.2 · p. 24, Table 11 · p. 25
  • Table 12 reports welfare realization under M0–M3 for the four-model subset in the contraction regime.Arrows indicate increases or decreases relative to M0 for the same model and regime.Source: §I.2 · p. 24, Table 12 · p. 25
Loading 2609.30940v1…