AgentMonopoly: Behavioral Patterns and Persona Control in LLM Agents

1 Introduction

• AgentMonopoly records messages, private explanations, executed actions, and outcomes to compare what agents say with what they do over a game.Purchases, trades, and rent payments also change the resources available for later turns and can affect opponents’ choices.Source: §1 · p. 2
• The study compares models under shared prompts, across sizes and opponent configurations, and in persona games against same-model opponents without the persona.Its questions concern behavioral patterns, performance by model size and opponents, and differences in persona agents’ words and actions.Source: §1 · pp. 1–2
• The tested models differ in trading, rent waiving, and cash and property management, while larger models do not consistently win more games.Some trading patterns appear at both tested sizes, but rent-waiving behavior can differ substantially within a model family.Source: §1 · p. 2
• Persona instructions change language and some actions, but an agent's decisions can still conflict with its stated priorities.Mistral’s Conservative agent declines a $400 purchase, then buys the property at auction for $420 while describing its approach as cautious.Source: §1 · p. 2

2 Related Work

• AgentMonopoly examines communication, trading, and rent waiving in repeated multiplayer games where agreements affect later ownership and affordability.Related work includes social simulations and games, collaboration and competition, and bargaining or resource-sharing tasks (Park et al., 2023; Vezhnevets et al., 2023; Fan et al., 2024; Lan et al., 2024; Zhou et al., 2024; Bianchi et al., 2024).Source: §2 · p. 2
• Unlike FinBoardBench (Hu et al., 2026), AgentMonopoly emphasizes trade communication, rent waiving, persona comparisons, and agreement between stated priorities and actions.FinBoardBench evaluates financial decision-making in Cashflow, Acquire, and Monopoly, focusing on financial consequences.Source: §2 · p. 2
• Prior persona studies examine trust games (Ou et al., 2025), strategic board-game move rules, or actions across investigation-game stories (Wang et al., 2026).AgentMonopoly keeps the game setting fixed and examines language and actions as opponents respond and earlier decisions affect later options.Source: §2 · pp. 2–3
• The study’s comparison with related work is organized around research questions and experimental design.Its analysis considers how interactions vary across models and how personas appear in words and actions during repeated competition.Source: §2 · p. 3

3 Research Questions

• RQ1 asks how models differ in trading, rent waiving, and cash and property management under the same prompt and rules, and which patterns recur across sizes.A behavioral profile summarizes actions measured under a particular game setup.Source: §3 · p. 3
• RQ2 asks how outcomes vary with model size and opponents, and whether win counts differ by turn-order position..Source: §3 · p. 3
• RQ3 asks how a persona agent’s language and actions differ from those of opponents without the persona.Persona control concerns how well instructions guide behavior; persona-associated words alone do not establish that decisions follow the instruction.Source: §3 · p. 3

4 Experimental Setup

• Experiments use four agents on a fixed 40-square board, with random events, property purchases, auctions, trades, building, rent waiving, and cash management.A player is eliminated when its cash becomes negative; games end when one player remains or at the 150-step limit, when highest net worth wins.Source: §4 · p. 3
• At each decision point, the engine supplies game information and allowed actions, then checks and records the agent’s proposed action, messages, and game changes.The shared prompt encourages bluffing and permits strategic rent waiving, while prohibiting public messages that contradict the chosen action.Source: §4 · p. 3
• The study tests Qwen3, DeepSeek-R1, Gemma, and Mistral at approximately 24–32 billion and 8–12 billion parameters, with five seeds per configuration.Models use publicly available weights and run locally at temperature 0.0; parameter counts do not imply equal capabilities across families.Source: §4 · p. 3
• Same-model games use four copies of one model version, while mixed-size games pair two larger and two smaller models from the same family.Cross-family games place one model from each family in a size group and rotate turn order; the configurations use five seeds per arrangement.Source: §4 · pp. 3–4
• Persona games assign one larger model at P1 a persona, while three same-model opponents use the base prompt; each family receives a different persona.The personas ask Qwen to propose fewer trades, DeepSeek to rely less on others, Gemma to interact more, and Mistral to keep more cash; persona effects cannot be separated from model-specific responses.Source: §4 · p. 4
• Measurements cover trade proposals and acceptance, rent waiving, cash and property management, wins by size, family, and position, and persona-related language and actions.Persona comparisons use the agent’s three opponents, whose positions and responses can affect how the game unfolds.Source: §4 · p. 4

5.1 RQ1: Behavioral patterns under shared rules

• Trade, acceptance, and rent-waiver profiles varied: DeepSeek-R1-8B accepted the smallest share of offers among smaller models but waived rent most often in its family.Rent waiving also differed sharply between the two tested sizes within the Qwen, Mistral, and Gemma families.Source: §5.1 · pp. 4–5
• Across same-model games, Qwen proposed trades most often, Mistral accepted the largest share of offers, and Gemma proposed the fewest at both sizes.These patterns do not establish a fixed strategy for a model family or show that behavior persists against other opponents.Source: §5.1 · p. 5, Figure 2 · p. 5

5.2 RQ2: Model size, opponents, and game outcomes

• 18 larger-model wins versus 22 smaller-model wins across 40 mixed-size games; larger models did not consistently win more across families.Qwen favored larger models 7–3, while DeepSeek, Mistral, and Gemma each favored smaller models; these counts describe the tested games.Source: §5.2 · pp. 5–6
• Cross-family games preserved the ranking of trade-proposal rates across families, while acceptance varied by configuration; Gemma4-26B acceptance rose from 17% to 56%.The comparison is between same-model and cross-family games.Source: §5.2 · p. 6
• Across 80 games, P1–P2 had 41 wins and P3–P4 had 39, a near balance that cannot rule out position effects in particular configurations.The 80 games combine different experiments, and the persona agent always played at P1 in persona games.Source: §5.2 · p. 6, Figure 3 · p. 6

5.3 RQ3: Language and actions under personas

• Persona-related language and message length changed, with message-length shifts differing by role: Zen messages averaged 68 characters, versus 108 for its opponents.Diplomatic messages averaged 122 characters versus 75 for opponents; Conservative messages averaged 109 versus 91, while Independent messages changed little.Source: §5.3 · p. 6, Figure 4 · p. 7
• Persona-linked actions were mixed: Zen traded less while continuing to build, Independent continued trading, Diplomatic showed no clear evidence of alliances, and Conservative kept more cash while accepting most offers.The reported action patterns differ by persona and do not establish a fixed strategy that overrides instructions.Source: §5.3 · pp. 6–7
• Because personas used different wording and were assigned to different models, the study cannot determine whether more specific wording explains observed changes.The authors propose assigning each persona to each model and testing decisions in identical game situations to separate instruction effects from opponents’ responses and events.Source: §5.3 · p. 7

6 Conclusion

• Persona instructions changed agents’ words and some actions, but not consistently across every behavior associated with a role; assessing instruction-following requires examining communication, actions, and outcomes together.The game engine and recorded interactions support further comparisons with other prompts, personas, model families, and player counts.Source: §6 · pp. 7–8

Scope of the environment

• The experiments cover four-agent games on one modified Monopoly board, whose optional rent waiving, bankruptcy rule, step limit, and winner criterion may limit transfer to other settings.The platform supports up to six agents, but player-count effects were not tested; renaming streets does not rule out Monopoly familiarity from model training.Source: Limitations · p. 8

Models, game counts, and uncertainty

• Five seeds per configuration, fixed temperature, and a limited set of public-weight model versions constrain how broadly the observed patterns can be generalized.Many logged actions are not independent games, and size comparisons cannot separate parameter count from training or design differences.Source: Limitations · p. 8

Dependence on instructions and opponents

• Persona findings depend on the shared prompt and opponents’ responses, and each family receives only one persona selected after baseline behavior was examined.The experiments do not determine whether more specific instructions work better; opponents are not an unaffected comparison group, and personas are not tested across every model family.Source: Limitations · pp. 8–9

Interpreting the measures

• Generated explanations and persona-related word counts do not establish intent or sustained role-following, while the hardship flag does not establish dishonesty.An agent flagged for claiming hardship may still face large upcoming payments, and waiving rent may help it win rather than reflect concern for others.Source: Limitations · p. 9

• The experiments use simulated games without human participants or real financial transactions, so game behavior does not establish human-like motives, moral character, safety, or trustworthiness.The studied behaviors include deceptive messages, threats, and eliminating opponents within the game.Source: Ethical Considerations · p. 9
• The environment could be used to develop more effective deception or secret coordination, but the experiments neither recommend bluffing in practice nor show how to prevent manipulation or collusion.Connecting agents to people, external tools, or real money and assets requires a separate risk assessment and appropriate safeguards.Source: Ethical Considerations · p. 9
• Research materials should include prompts, rule changes, model identifiers, and limitations, with generated transcripts labeled and personal information or harmful content reviewed before release.Reuse of models and software should follow their licenses; these practices support reproducibility and reduce misleading claims or inappropriate use.Source: Ethical Considerations · p. 9

A.1 Game Mechanics and logging

• The engine gives the acting agent its financial and positional state, opponent and property information, and decision options, then records its proposed action and two text fields.The fields are a private generated explanation and a public message; logged data are organized as player states, board states, and recent interaction history.Source: §A.1 · p. 10
• A game step includes a dice roll and associated decisions and events, with procedures for jail, property management, movement, landing resolution, doubles, auctions, invalid trades, and invalid outputs.For the net-worth winner criterion, net worth is cash plus half the purchase price of owned properties and half the construction cost of buildings.Source: §A.1 · p. 10
• The platform includes an interactive replay and analysis mode that visualizes cash-flow trends.The text illustrates the mode with Fig. 7.Source: §A.1 · p. 10

A.2 Shared prompt and persona instructions

• The common prompt describes agents as cunning and autonomous, encourages outsmarting opponents and bluffing, and prohibits public messages that contradict actions.It permits strategic rent waiving, including to keep a weaker opponent in the game against a stronger rival.Source: §A.2 · p. 10
• Persona experiments supplement the common base prompt with an additional instruction for the designated agent, with four configurations intended to change behavior.The text directs readers to Table 2 for the configurations and intended changes.Source: §A.2 · p. 10

A.3 Experimental configurations

• Table 3 summarizes four experimental series and their game counts; each condition uses five seeds at temperature 0.0, with models drawn from four families at two sizes each.The model identifiers are listed by family and relative size in Table 4.Source: §A.3 · pp. 10–11, Table 3 · p. 12, Table 4 · p. 12
• Seating is balanced by rotating model sizes across P1–P4 in S2 and using four cyclic turn orders in S3 so each family takes the first position once.Table 2 specifies persona conditions: only P1 receives the persona, P2–P4 use the same model with the base prompt, and each row uses seeds 42–46.Source: §A.3 · p. 11, Table 2 · p. 11

A.4 Behavioral and language measures

• Trading measures include proposals per 100 dice rolls, offer acceptance, and the fraction of rent decisions in which the owner waives payment.Source: §A.4 · p. 11
• Property measures cover holdings of purchasable spaces and buildings, with houses counted individually and each hotel counted as five houses.“Houses at the end” records final holdings at the final game step, at most step 150; family averages divide final house counts by repeated games.Source: §A.4 · p. 11
• Language measures report the fraction of public messages or private explanations containing a persona-related term and characters per public message.Source: §A.4 · p. 11

B.1 Win distribution in mixed-size games

• 22 wins for smaller models versus 18 for larger models overall; the family-level advantage varied, and no player position consistently dominated.Larger models won more in Qwen, while smaller models won more in the other three families.Source: §B.1 · p. 12, Figure 8 · p. 13

B.2 Game outcomes in corss-family games

• Trade-proposal-rate rankings remained unchanged across cross-family and same-model games in both size groups, while acceptance rates showed model-specific differences.Table 5 reports cross-family results with corresponding same-model measures in parentheses; each model played 20 games across five positions.Source: §B.2 · p. 12, Table 5 · p. 13