A reproducible framework for studying emergent economic behavior with LLM agents in synthetic marketplaces

A research framework that treats LLMs as bounded policy approximators in a controlled marketplace environment. Study how cognitive architectures shape economic outcomes through systematic experimentation with freelancer and client agents.

• 🧠 Strategic Decision-Making: LLM agents develop bidding strategies through natural language reasoning
• 📈 Market Dynamics: Track competition, efficiency, and inequality patterns over time
• 🎭 Behavioral Adaptation: Agents learn and adapt through reflection mechanisms
• 📊 Controlled Experiments: Systematic comparison of agent configurations (LLM vs Random vs Hybrid)
• 🔬 Reproducible Results: Complete framework available as open-source software

• Python 3.8+
• OpenAI API key or custom LLM endpoint

git clone [REPOSITORY_URL]
cd simulated_marketplace
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate
pip install -r requirements.txt

# Option A: Environment variable
export OPENAI_API_KEY='sk-your-key-here'

# Option B: Configuration file
cp config/private_config.example.py config/private_config.py
# Edit with your API credentials

# Test your connection
python test_llm_connection.py

# Run basic simulation
python run_marketplace.py --freelancers 20 --clients 5 --rounds 50

# Analyze results
python analyze_marketplace.py --simulation-file results/your_simulation.json

Simulation Engine

• TrueGPTMarketplace: Main simulation with LLM-powered agents
• SimpleReputationManager: Multi-tier reputation system (New → Established → Expert → Elite)
• JobRankingCalculator: Semantic skill matching and job scoring

Agent Types

• LLM Agents: Strategic reasoning through structured prompts
• Random Agents: Probabilistic baseline (50% bid chance)
• Greedy Agents: Simple rational heuristics (highest budget preference)

Analysis Tools

• MarketplaceAnalysis: Comprehensive metrics and visualizations
• Statistical validation with confidence intervals
• Temporal trend analysis and adaptation tracking

✅ Controlled Experimentation

• Fixed population size for systematic comparison
• Configurable parameters (bid limits, reflection rates, cooldown periods)
• Multiple baseline agents for rigorous benchmarking

✅ Advanced Reputation System

• Performance-based tier progression
• Historical reputation tracking across rounds
• Reputation-aware decision making in agent prompts

✅ Market Mechanics

• Fixed-budget bidding (no rate negotiation)
• Natural skill distribution across job categories
• Bid cooloff system enabling re-bidding after N rounds
• Dynamic budget adjustments by clients

✅ Type Safety & Validation

• Pydantic models for all LLM responses (8 distinct types)
• Comprehensive error handling and logging
• 174+ automated tests with baseline agent validation

# Compare agent reasoning capabilities
python run_marketplace.py --baseline-scenario random     # Noise baseline
python run_marketplace.py --baseline-scenario greedy     # Rational baseline  
python run_marketplace.py                                # Full LLM agents

# Study reflection mechanisms
python run_marketplace.py --reflection-probability 0.0   # No reflections
python run_marketplace.py --reflection-probability 0.1   # Low reflection rate

# Bid cooloff effects
python run_marketplace.py --bid-cooloff-rounds 0    # No re-bidding
python run_marketplace.py --bid-cooloff-rounds 5    # 5-round cooloff

# Job posting frequency
python run_marketplace.py --job-posting-cooldown-min 1 --job-posting-cooldown-max 3   # High frequency
python run_marketplace.py --job-posting-cooldown-min 5 --job-posting-cooldown-max 15  # Low frequency

# Large-scale experiments
python run_marketplace.py --freelancers 200 --clients 30 --rounds 100 --quiet --max-workers 20

# Accelerated progression
python run_marketplace.py --max-active-jobs 5  # Higher freelancer capacity

The framework makes several key assumptions to enable controlled experimentation:

Market Structure

• Fixed agent population (no entry/exit dynamics)
• Discrete time rounds with synchronous actions
• Binary hiring decisions (one freelancer per job)

Agent Behavior

• LLMs as bounded policy approximators with natural language reasoning
• Perfect memory through reflection system
• Reputation-aware decision making

Economic Model

• Fixed budgets with transparent pricing
• No transaction costs or platform fees
• Automatic job completion (success = True) to avoid evaluation bias

Technical Implementation

• Category-first job generation (business domains assigned to clients)
• Natural skill distribution (organic clustering around popular fields)
• Pure ranking system (no arbitrary threshold cutoffs)

See FRAMEWORK_ASSUMPTIONS.md for complete details

All results are saved to results/ directory:

• Raw Data: Complete interaction logs in JSON format
• Analysis Reports: Statistical summaries with confidence intervals
• Visualizations: Publication-ready figures (market trends, agent learning, comparative analysis)
• Reputation Tracking: Historical progression data for longitudinal studies

Comprehensive testing ensures framework reliability:

# Run full test suite (174+ tests)
python -m pytest

# Test specific components
python -m pytest tests/test_baseline_agents.py      # Baseline agent validation
python -m pytest tests/test_marketplace_integration.py  # Integration tests

Immediate Extensions

• Cross-model validation (different LLM architectures)
• Human-AI hybrid markets
• Alternative auction mechanisms

Long-term Opportunities

• Real-world platform validation
• Policy implications research
• Large-scale economic ecosystem studies

This research demonstrates AI-driven scientific discovery. Contributions welcome for:

• Extended research methodologies
• Technical improvements and optimizations
• Validation studies and cross-platform testing
• Documentation and tutorial improvements

See CONTRIBUTING.md for detailed guidelines.

Silvia Terragni
Email: silviaterragni at upwork.com

If you use this framework in your research, please cite our paper:

@inproceedings{terragni2025simulating,
  title={Simulating Two-Sided Job Marketplaces with AI Agents},
  author={Terragni, Silvia and Nojavanasghari, Behnaz and Yang, Frank and Rabinovich, Andrew},
  booktitle={Open Conference of AI Agents for Science},
  year={2025},
  url={https://openreview.net/forum?id=pjpkEHH5YS}
}

Paper: Simulating Two-Sided Job Marketplaces with AI Agents

This research was conducted almost entirely by AI agents as part of the Agents4Science 2025 conference exploring AI-generated scientific research. The work represents a novel approach to scientific inquiry where artificial intelligence serves as both researcher and subject.