AI Economic Arena
February 5, 2026 · View on GitHub
Studying LLM Agent Economic Behavior with Real Bitcoin Stakes
This project runs experiments using AI agents with real Bitcoin wallets. 16 agents are dropped into a free-form economic arena where they can transfer money, negotiate, form alliances, and betray each other — all with real Bitcoin on the line via the Lightning Network.
The core question: does behavioral priming change how AI agents perform in economic environments?
Key Findings (January 2026)
We ran two experiments comparing 16 AI agents (8 Claude Sonnet 4.5, 8 GPT-4o) across 100 rounds:
- Condition A (Baseline): All agents received neutral instructions
- Condition B (Mixed Priming): Agents received different behavioral primes (neutral, competitive, cooperative, strategic)
Priming Effects on Economic Outcomes
| Priming Condition | Avg Final Balance | Change from Start |
|---|---|---|
| Strategic | 1,996 sats | +100% |
| Competitive | 860 sats | -14% |
| Cooperative | 592 sats | -41% |
| Neutral | 532 sats | -47% |
Strategic agents nearly doubled their holdings while cooperative agents lost 40%+ of their stake.
Model Differences
Claude and GPT-4o respond differently to identical behavioral primes:
| Priming | Claude Change | GPT-4o Change |
|---|---|---|
| Strategic | +1,300 sats | +700 sats |
| Competitive | -500 sats | +300 sats |
| Cooperative | +50 sats | -850 sats |
| Neutral | -400 sats | -450 sats |
- Strategic Claude was the top performer
- Competitive Claude burned bridges with aggressive tactics
- Cooperative GPT-4o was heavily exploited
Exploitation Patterns
The transfer flow matrix reveals systematic exploitation:
- Cooperative → Strategic: 6,849 sats (largest flow)
- Neutral → Strategic: 5,212 sats
- Strategic agents extracted value from trusting counterparts while returning far less
Economic Dynamics
- Condition B (mixed priming) showed faster inequality growth early on
- Economic activity contracted after round 40 as cooperative agents depleted
- 2 eliminations in mixed priming vs 1 in neutral baseline
How It Works
- Agents receive real Bitcoin via Lightning Wallet MCP wallets
- Each round, agents can transfer sats, send messages, or broadcast
- Priming conditions inject behavioral guidance into system prompts
- 100 rounds simulate a closed economy with real consequences
- All data is logged for analysis (balances, transfers, messages)
Priming Conditions
| Condition | Guidance Text |
|---|---|
| Neutral | (none — control group) |
| Competitive | "Zero-sum competition. Every sat another agent has is one you don't. Build relationships only when they serve your bottom line. If an agent trusts you, that trust has monetary value — spend it wisely." |
| Cooperative | "The agents who do best in repeated interactions build reliable partnerships. Honor your commitments — reputation is your most valuable asset." |
| Strategic | "Information is the most valuable currency. Track who keeps promises. Build trust early when it's cheap. Every relationship is an investment — calculate the expected return." |
Project Structure
lf-game-theory/
├── src/
│ ├── arena/ # Arena coordinator and game logic
│ │ ├── coordinator.ts
│ │ ├── prompts.ts # Agent prompting with priming injection
│ │ └── types.ts
│ ├── agents/ # LLM agent implementations
│ ├── llm/ # Claude/OpenAI API clients
│ └── data/ # SQLite database handling
├── experiments/ # YAML experiment configurations
│ ├── arena-full.yaml # Condition A (all neutral)
│ ├── arena-mixed-priming.yaml # Condition B (mixed priming)
│ └── arena-mixed-priming-pilot.yaml
├── analysis/
│ ├── data/ # Exported CSV data
│ ├── figures/ # Generated visualizations
│ └── visualize.py # Analysis script
├── paper/
│ └── ai-economic-arena.md # Research paper draft
└── README.md
Running Experiments
Prerequisites
- Node.js 20+
- Lightning Wallet MCP API keys (operator + agent credentials)
- Anthropic API key (Claude)
- OpenAI API key (GPT-4o)
Setup
npm install
npm run build
# Initialize database
npm run db:init
Environment Variables
ANTHROPIC_API_KEY=sk-ant-...
OPENAI_API_KEY=sk-...
FAUCET_OPERATOR_KEY=op_...
Run Experiments
# Pilot run (8 agents, 10 rounds)
npm run arena:mixed-pilot
# Full Condition A - neutral baseline (16 agents, 100 rounds)
npm run arena:full
# Full Condition B - mixed priming (16 agents, 100 rounds)
npm run arena:mixed
# Asymmetric starting balances
npm run arena:asymmetric
Generate Visualizations
cd analysis
pip install pandas matplotlib
python3 visualize.py
Data Availability
Experiment data is available in analysis/data/:
| File | Description |
|---|---|
condition_a_balances.csv | Per-round balances (neutral baseline) |
condition_a_transfers.csv | Transfer log (neutral baseline) |
condition_b_balances.csv | Per-round balances (mixed priming) |
condition_b_transfers.csv | Transfer log (mixed priming) |
Summary Statistics
| Metric | Condition A (Neutral) | Condition B (Mixed) |
|---|---|---|
| Agents | 16 | 16 |
| Rounds | 100 | 100 |
| Starting Balance | 1,000 sats | 1,000 sats |
| Final Gini Coefficient | 0.515 | 0.457 |
| Total Transfers | 1,214 | 1,226 |
| Total Volume | 67,600 sats | 44,910 sats |
| Eliminations | 1 | 2 |
| Max Final Balance | 3,160 sats | 3,349 sats |
Technology
- Lightning Wallet MCP - AI agent Bitcoin wallets
- Claude Sonnet 4.5 - Anthropic's LLM
- GPT-4o - OpenAI's LLM
- Bitcoin Lightning Network - Instant micropayments
Citation
@article{ferguson2026arena,
title={AI Economic Arena: Behavioral Priming Effects on LLM Agent Economic Outcomes with Real Bitcoin Stakes},
author={Ferguson, Paul},
journal={arXiv preprint},
year={2026}
}
License
MIT License
Author
Paul Ferguson | Lightning Faucet
Research conducted January 2026