The Automated but Risky Game Modeling and Benchmarking agent-to-agent negotiations and transactions in consumer markets.

Shenzhe Zhu1, Jiao Sun2, Yi Nian3, Tobin South4, Alex Pentland4,5, Jiaxin Pei5,†

1University of Toronto · 2Google DeepMind · 3University of Southern California · 4MIT · 5Stanford University · Corresponding Author

Contact: cho.zhu@mail.utoronto.ca · pedropei@stanford.edu

A2A Negotiation Arena overview showing buyer and seller agents negotiating consumer transactions
AI agents negotiate prices, accept or reject offers, and expose measurable transaction risks.

Overview

What happens when both sides of a consumer transaction are automated?

A2A Negotiation Arena studies consumer-market deals where both buyers and sellers delegate negotiation to AI agents. We evaluate how different LLM agents perform as deal-makers, and how automation can create imbalanced games, overspending, invalid deals, or stalled negotiations.

Framework

Negotiation traces become measurable transaction outcomes.

Both agents know the retail price, while only the seller knows wholesale cost. The buyer operates under a budget limit, the seller should not go below wholesale, and multi-turn offers expose capability gaps and anomaly patterns.

Agent-to-agent negotiation and transaction framework

Key risks

Automation changes who wins, who loses, and how failures appear.

Risk 1

Imbalanced game

When agents with different negotiation capacities compete, users represented by weaker agents can face strategic disadvantage and economic loss.

Imbalanced game risk illustration
Risk 2

Anomaly Behavior & Outcome

Even when agents reach a deal, anomaly behavior can violate constraints, omit required fees, substitute products, refuse feasible offers, or stall negotiations.

Leaderboard

Ranked negotiation and risk performance

Relative Profit

Higher is better

Seller-side average profit from clean deals, normalized against the configured 1.0x baseline.

Rank Model Relative Profit Seller PRR Buyer PRR

Resources

Reproduce and extend the benchmark

The repository includes experiment runners, product datasets, anomaly-labeling scripts, and analysis notebooks for reproducing the benchmark.

Citation

BibTeX

@misc{zhu2025automatedriskygamemodeling,
      title={The Automated but Risky Game: Modeling and Benchmarking 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},
}