Advancing AI Negotiations: A Large-Scale Autonomous Negotiation Competition

TL;DR

This work investigates AI negotiation by embedding classic negotiation theory within autonomous AI agents and testing their performance at scale using an Axelrod-inspired round-robin competition. The study demonstrates that warmth, an interpersonal style cue, consistently improves both objective outcomes (value claimed, value created) and counterpart perceptions, while dominance enhances value claiming within deals but increases impasses. It also uncovers AI-specific strategies—such as chain-of-thought reasoning and prompt injection—that can substantially affect negotiation outcomes beyond traditional tactics, suggesting a need for a unified theory of AI negotiation that integrates human realism with AI-enabled mechanisms. The findings have practical implications for designing robust AI negotiators and for understanding how social signals and AI-specific techniques jointly shape autonomous negotiations, while outlining important directions for extending analysis to repeated interactions, human-AI contexts, and broader model architectures.

Abstract

We conducted an International AI Negotiation Competition in which participants designed and refined prompts for AI negotiation agents. We then facilitated over 180,000 negotiations between these agents across multiple scenarios with diverse characteristics and objectives. Our findings revealed that principles from human negotiation theory remain crucial even in AI-AI contexts. Surprisingly, warmth -- a traditionally human relationship-building trait -- was consistently associated with superior outcomes across all key performance metrics. Dominant agents, meanwhile, were especially effective at claiming value. Our analysis also revealed unique dynamics in AI-AI negotiations not fully explained by existing theory, including AI-specific technical strategies like chain-of-thought reasoning and prompt injection. When we applied natural language processing (NLP) methods to the full transcripts of all negotiations, we found positivity, gratitude, and question-asking (associated with warmth) were strongly associated with reaching deals as well as objective and subjective value, whereas conversation lengths (associated with dominance) were strongly associated with impasses. The results suggest the need to establish a new theory of AI negotiation, which integrates classic negotiation theory with AI-specific negotiation theories to better understand autonomous negotiations and optimize agent performance.

Figures (49)

• Figure 2: Agent warmth and dominance shape objective and subjective negotiation outcomes. (A-E) Standardized regression coefficients with 95% confidence intervals for warmth (red circles) and dominance (blue triangles) across negotiation outcomes. Coefficients are standardized to enable direct comparison of effect sizes across outcomes and contexts. (F-I) Response surfaces showing the relationships between agent warmth($x$-axis) and dominance ($y$-axis) combinations and specific outcomes: (F) Points earned by agent (in integrative negotiations), (G) Value claimed by agent (in distributive negotiations), (H) Value created (in integrative negotiations), (I) Counterpart satisfaction ratings (in distributive and integrative negotiations). Contours generated using inverse distance weighting interpolation with $k$-means clustering for optimal bin placement.
• Figure 3: Warmth and dominance effects differ when conditioning on whether the agent reaches a deal. (A-G) Standardized regression coefficients with 95% confidence intervals for warmth (red circles) and dominance (blue triangles). Gray boxes on the left show the rate of reaching a deal across all negotiations. White boxes on the right show outcomes conditional on reaching a deal. Coefficients are standardized to enable direct comparison of effect sizes across outcomes and contexts. (H-K) Response surfaces showing the relationships between agent warmth ($x$-axis) and dominance ($y$-axis) combinations and specific outcomes: (H) Rate of reaching a deal, (I) Points earned when deal is reached, (J) Value claimed when deal is reached, (K) Value created when deal is reached. Contours generated using inverse distance weighting interpolation with $k$-means clustering for optimal bin placement.
• Figure 4: Warmth and dominance shape communication strategies that drive negotiation success. (A-H) Standardized regression coefficients with 95% confidence intervals for warmth (red circles) and dominance (blue triangles) effects on communication characteristics during negotiations. (I-M) Standardized regression coefficients with 95% confidence intervals showing how eight communication strategies influence negotiation outcomes across different measures. Coefficients are standardized to enable direct comparison of effect sizes across outcomes and contexts.
• Figure : Competition Overview and Participant Demographics
• Figure S1: Competition instructions provided to participants in the preliminary round.