Cooperate to Compete: Strategic Coordination in Multi-Agent Conquest

2025-11-19 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Data Science & Analytics · Depth: Expert, extended

Summary

Researchers from the University of California, Berkeley, introduce "Cooperate to Compete" (C2C), a multi-agent environment designed to test Language Model (LM)-based agents in mixed-motive settings requiring short-term cooperation for long-term competitive goals. C2C is a long-horizon game where players have asymmetric, secret objectives to conquer regions on a map, engage in private, non-binding negotiations, and can form and break alliances. The study involved over 1,100 games, including AI-only matches and a user study pitting human players against AI opponents using models like Gemini 3.1 Pro, Grok 4.1, and GPT 5.2. Key findings indicate humans are more aggressive negotiators, accepting deals without counteroffers only 56.3% of the time compared to 67.6% for LMs, and are less reliable partners. Targeted prompting interventions, such as encouraging aggressive negotiation, support seeking, and deception, improved LM agent win rates from 22.2% to 32.7%.

Key takeaway

For NLP Engineers developing multi-agent systems, this research highlights that LMs can be engineered to navigate complex, mixed-motive environments effectively. You should consider implementing targeted prompting strategies to refine negotiation behaviors, such as encouraging aggressive deal-making or strategic deception, to improve agent performance in competitive scenarios. This approach can lead to more robust and adaptable AI agents capable of sophisticated coordination.

Key insights

LM agents can be prompted to improve strategic coordination and negotiation in complex, mixed-motive environments.

Principles

• Strategic coordination is critical for success in mixed-motive environments.
• Non-binding negotiations enable dynamic alliance formation and dissolution.
• Prompting can significantly alter LM agent negotiation behaviors and performance.

Method

The C2C environment uses a modified Risk-like game with fog-of-war, secret objectives, and non-binding natural language negotiations to study multi-agent strategic coordination.

In practice

• Implement prompt-based interventions to enhance LM agent negotiation tactics.
• Design multi-agent environments with asymmetric information and non-binding agreements.
• Analyze human-AI behavioral differences to inform agent strategy development.

Topics

• Multi-Agent Systems
• Language Model Agents
• Strategic Negotiation
• Mixed-Motive Games
• Human-AI Interaction

Code references

• abbykoneill/negotiationgames

Best for: NLP Engineer, AI Scientist, Research Scientist, AI Engineer

Related on AIssential

• Talk is Cheap, Communication is Hard: Dynamic Grounding Failures and Repair in Multi-Agent Negotiation — Dynamic grounding, not just static understanding, is critical for effective multi-agent LLM coordination and negotiation.
• Benchmarking Emergent Coordination in Large-Scale LLM Populations: An Evaluation Framework on the MoltBook Archive — Autonomous LLM agents exhibit emergent structural roles and information diffusion patterns, but struggle with effective decentralized cooperation.
• Designing Effective Multi-Agent Architectures — Effective multi-agent system performance is an architectural outcome, not solely a prompting problem.
• Skill-MAS: Evolving Meta-Skill for Automatic Multi-Agent Systems — Skill-MAS decouples experience retention from LLM parametric updates using an evolvable Meta-Skill for multi-agent systems.
• Interview with Xinwei Song: strategic interactions in networked multi-agent systems — Strategic interactions in networked multi-agent systems can be managed through mechanism design and reputation-aware reinforcement learning.
• Multi-Agent Prompt Engineering is Dead: NEW Proof Your Agents Aren't Talking — Multi-agent LLM prompt optimization is often unnecessary; focus on independent agent tuning and architectural harnesses.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.