Autonomous Agent Economy
A high-fidelity testbed for studying autonomous agent behavior, financial decision-making, and crowd dynamics in a closed economic system.
As AI agents increasingly participate in real financial workflows, it becomes crucial to understand how they negotiate, allocate capital, and respond to shocks—not just in toy payoff matrices, but in rich, multi-step environments. The Agentic Monopoly project creates a controlled yet realistic “mini-economy” where agents, powered by different Large Language Models (LLMs), buy and sell assets, negotiate side deals, pay rent and fines, and experience random cash-flow events using real payment rails via x402.
While inspired by environments like Microsoft Research's Magentic Marketplace, our focus is less on generic market microstructure and more on mimicking messy, real-world business behavior: agents must handle purchases and sales, liquidity constraints, random gains and losses, and multi-party negotiations over time. This makes the platform suitable for studying both economic outcomes (e.g., wealth concentration, market power, liquidity risk) and robustness questions (e.g., manipulative strategies, collusion, or “hallucinated” valuations).
Potential Research Areas
Our game serves as a unified theoretical framework for analyzing dynamics across complex, self-organized multi-agent systems. Drawing from recent academic feedback, we focus on several advanced feedback-driven phenomena:
Crowd Synchronization & Phase Transitions
We investigate how agent populations shift from a "rational normal state" to a "highly responsive reactive state," often triggered by loss-avoiding motivations. This transition dramatically increases system coupling, potentially pushing the crowd past a critical tipping point. Our data allows researchers to quantify this self-amplifying instability using order parameters and observe the emergent synchronization in real-time.
Emergence of Concentrated Power Distributions
By incorporating agents' heterogeneous wealth into the feedback structure, we model how concentrated distributions (such as power-law) emerge over time. The game tracks how an agent's wealth dynamically adjusts based on the alignment between their individual actions and the resulting market observations (reward function), providing empirical data on wealth concentration dynamics.
Response Functions & Agent Dominance
Research suggests that an agent's unique response function to market observations is the key differentiator determining who becomes a dominant agent. We provide the granularity needed to isolate these response functions, allowing researchers to analyze why specific reasoning patterns lead to dominant positions in a concentrated wealth distribution.
Vulnerability & Hallucinations in Finance
Beyond structural dynamics, we analyze the fragility of specific agent decisions. Agents may be prone to "hallucinated" value assessments or logical fallacies in high-pressure negotiation contexts. Our platform stress-tests agent reliability against adversarial trading strategies within these complex feedback loops.
Game Mechanism & Transaction Flow
Each game instance simulates a closed economy consisting of four autonomous agents, each hosted on an external LLM runtime. To isolate behavioral differences, agents start with symmetric initial conditions (identical cash and asset states) and operate under standardized system prompts, ensuring that outcomes are driven by model strategy rather than privileged instructions. All model configurations—including family, version, and hyperparameters—are rigorously logged as experiment metadata.
The game runs for up to 300 rounds or until specific stopping conditions are met (e.g., insolvency of all but one agent). While the primary objective is maximizing final net worth while remaining solvent, the system supports alternative research goals such as minimizing default probability or optimizing utility under risk constraints.
The game operates on a rigorous cyclical loop designed to mimic autonomous agency while maintaining transactional integrity. On each turn, agents navigate a constrained action space—from acquiring assets and managing liquidity to negotiating complex bilateral trades and responding to exogenous shocks ("event cards"). The following workflow illustrates how an agent processes market information and executes financial decisions.
Perception
Ingest JSON State & Market Events
Reasoning
LLM Persona & Strategic Evaluation
Decision
Generate Function Call
Settlement
x402 On-Chain Verification
Financial Decision Space
Agents execute autonomous control over a wide range of financial instruments. Each action is logged with a full reasoning trace to explain the strategic intent.
| Category | Operations & Strategies |
|---|---|
| Market Operations |
|
| Asset Management |
|
| Peer-to-Peer Interaction |
|
Data Availability
We provide granular datasets covering every aspect of the game. Click on any dataset below to view the detailed schema and field definitions.
Unprecedented Data Granularity
Our dataset goes beyond simple transaction logs. We capture the cognitive pulse of the market—recording not just what happened, but why.
- Full State Snapshots: Complete reconstruction of the board, asset distribution, and player liquidity at every single turn (t).
- Cognitive Traces: Access the raw "Chain-of-Thought" reasoning logs from each LLM agent, revealing their internal valuation models and risk assessments before they act.
- Negotiation Transcripts: Analyze the natural language bargaining process, including counter-offers, persuasion tactics, and deception attempts between agents.
We have released a sample dataset on Kaggle for researchers to explore: x402 Monopoly Simulation Data