The Automated but Risky Game: Modeling Agent-to-Agent Negotiations and Transactions in Consumer Markets
Shenzhe Zhu
(
- 2025/05/17: We have released our code and dataset.
This repository contains the implementation of an automated negotiation system that simulates agent-to-agent negotiations in consumer markets. The system uses large language models (LLMs) to power both buyer and seller agents, enabling realistic and dynamic price negotiations. We also provide methods for detecting model anomalies and potential risks in automated negotiations.
- Create a conda environment:
conda create -n negotiation python=3.9
conda activate negotiation- Install dependencies:
pip install -r requirements.txt- Set up API keys in a
Config.pyfile:
OPENAI_API_KEY = "your_openai_api_key"
DEEPSEEK_API_KEY = ["your_deepseek_api_key1", "your_deepseek_api_key2"]
ZHI_API_KEY = ["your_zhizengzeng_api_key1", "your_zhizengzeng_api_key2"]
GOOGLE_API_KEY = "your_google_api_key"
Run experiments using the provided shell script:
./run_all.shOr run individual experiments using main.py:
python main.py \
--products-file dataset/products.json \
--buyer-model gpt-3.5-turbo \
--seller-model gpt-3.5-turbo \
--summary-model gpt-3.5-turbo \
--max-turns 30 \
--num-experiments 1 \
--output-dir resultsThe system tests five different budget scenarios for each product:
- High: Retail Price * 1.2
- Retail: Retail Price
- Mid: (Retail Price + Wholesale Price) / 2
- Wholesale: Wholesale Price
- Low: Wholesale Price * 0.8
- OpenAI
- DeepSeek
- Qwen
Results are saved in the results directory with the following structure:
results/
└── seller_{seller_model}/
└── {buyer_model}/
└── product_{product_id}/
└── budget_{scenario}/
└── product_{product_id}_exp_{experiment_num}.json
Each result file contains:
- Complete conversation history
- Price offers
- Negotiation outcome
- Budget scenario
- Model information
In data_postprocess/draw_result.ipynb, we provide methods for calculating various metrics and generating visualizations, including:
- Price Reduction Rate
- Total Profit
- Deal Rate
- Profit Rate
We provide comprehensive model anomaly analysis tools in data_postprocess/draw_risk.ipynb, which includes methods for analyzing various types of model anomalies:
- Overpayment: Cases where the buyer pays significantly more than the market value
- Constraint Violation: Instances where negotiation constraints are not properly followed
- Deadlock: Situations where negotiations reach an impasse
.
├── main.py # Main experiment runner
├── Conversation.py # Conversation management and negotiation logic
├── LanguageModel.py # LLM interface and API handling
├── run_all.sh # Shell script for running multiple experiments
├── dataset/ # Contains product information
│ ├── products.json
│ └── products_mini.json
└── data_postprocess/ # Data processing and analysis tools
├── draw_result.ipynb # Calculate metrics and generate visualizations
└── draw_risk.ipynb # Model anomaly analysis
🧾 Citation If you find our work useful in your research or applications, please consider citing:
BibTeX:
@misc{zhu2025automatedriskygamemodeling,
title={The Automated but Risky Game: Modeling Agent-to-Agent Negotiations and Transactions in Consumer Markets},
author={Shenzhe Zhu and Jiao Sun and Yi Nian and Tobin South and Alex Pentland and Jiaxin Pei},
year={2025},
eprint={2506.00073},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2506.00073},
}