Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

NegotiationGym: Self-Optimizing Agents in a Multi-Agent Social Simulation Environment

Shashank Mangla, Chris Hokamp, Jack Boylan, Demian Gholipour Ghalandari, Yuuv Jauhari, Lauren Cassidy, Oisin Duffy

TL;DR

NegotiationGym addresses the need for a flexible, configurable platform to study multi-agent negotiation dynamics using LLM-based agents. It presents an open-source toolkit with a configuration-driven API, utility-aware agents, and a self-optimization loop where agents rewrite prompts based on cross-round feedback. The framework integrates AutoGen and a MongoDB-backed job queue, supporting CLI and GUI workflows and providing detailed outcome reports. A case study on buyer–seller negotiations with coaching demonstrates how reflective learning shifts utilities and reduces no-deal outcomes, illustrating the potential for autonomous strategy adaptation in social simulations.

Abstract

We design and implement NegotiationGym, an API and user interface for configuring and running multi-agent social simulations focused upon negotiation and cooperation. The NegotiationGym codebase offers a user-friendly, configuration-driven API that enables easy design and customization of simulation scenarios. Agent-level utility functions encode optimization criteria for each agent, and agents can self-optimize by conducting multiple interaction rounds with other agents, observing outcomes, and modifying their strategies for future rounds.

NegotiationGym: Self-Optimizing Agents in a Multi-Agent Social Simulation Environment

TL;DR

NegotiationGym addresses the need for a flexible, configurable platform to study multi-agent negotiation dynamics using LLM-based agents. It presents an open-source toolkit with a configuration-driven API, utility-aware agents, and a self-optimization loop where agents rewrite prompts based on cross-round feedback. The framework integrates AutoGen and a MongoDB-backed job queue, supporting CLI and GUI workflows and providing detailed outcome reports. A case study on buyer–seller negotiations with coaching demonstrates how reflective learning shifts utilities and reduces no-deal outcomes, illustrating the potential for autonomous strategy adaptation in social simulations.

Abstract

We design and implement NegotiationGym, an API and user interface for configuring and running multi-agent social simulations focused upon negotiation and cooperation. The NegotiationGym codebase offers a user-friendly, configuration-driven API that enables easy design and customization of simulation scenarios. Agent-level utility functions encode optimization criteria for each agent, and agents can self-optimize by conducting multiple interaction rounds with other agents, observing outcomes, and modifying their strategies for future rounds.

Paper Structure

This paper contains 14 sections, 2 equations, 5 figures.

Table of Contents

  1. Introduction
  2. Related Work
  3. NegotiationGym
  4. Configuration
  5. Running Simulations and Observing Outcomes
  6. Optimization
  7. Case Study: Buyer–Seller Negotiation Coaching
  8. Limitations
  9. Conclusion
  10. Acknowledgments
  11. Appendix
  12. Extended Related Work
  13. Simulation Configuration Example
  14. Reflection Prompt

Figures (5)

  • Figure 1: Cumulative average utility curves for each optimization mode on a 20-turn setting.
  • Figure 2: Average surplus shares over 20 negotiations for each optimization mode on a 20-turn setting (left) and a 10-turn setting (right). Points lie on or below the Pareto frontier. All prompts and configuration for this experiment are available in the open-source repository referenced above.
  • Figure 3: Simulation configuration example
  • Figure 4: Agent reflection prompt
  • Figure 5: Negotiation Coach Agent prompt