Advancing AI Negotiations: A Large-Scale Autonomous Negotiation Competition

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.

Introduction. Autonomous AI agents are transforming negotiations [1] and laying the groundwork for widespread agent-toagent negotiation across tasks and contexts at scale [2, 3]. Computer science research has used negotiation To bridge these fields and advance our understanding of AI negotiation, we conducted a large-scale, international AI negotiation competition in which participants designed and refined prompts for AI negotiation agents. Our methodology draws direct inspiration from Robert Axelrod’s seminal 1980s tournament approach to studying cooperation, which revolutionized game theory and evolutionary biology through its elegant competitive framework [26, 27] and motivated subsequent tournaments that also generated important findings about when collaborative versus competitive strategies prevail and how implicit coalitions emerge [28, 29].

Just as Axelrod invited experts to submit strategies for iterated Prisoner’s Dilemma Games—yielding profound insights about the emergence of cooperation that transcended disciplinary boundaries—our competition represents a similar opportunity to discover foundational principles of AI negotiation. By systematically pitting diverse negotiation strategies against one another in a round-robin format, we follow Axelrod’s template for uncovering fundamental principles that operate across contexts while adapting this approach to the unique challenges and opportunities of the AI era. In particular, we aim to advance recent work toward a unified theory of AI negotiation through discoveries at the intersection of classic human negotiation theory and technical AI negotiation theory by combining computer science approaches with behavioral and cognitive insights from human negotiations.

We conceptualize warmth and dominance following the Interpersonal Circumplex (IPC) framework [54, 55], which characterizes individual-level interpersonal behaviors along two orthogonal dimensions: warmth and dominance. Negotiation theory and empirical evidence similarly suggests that negotiators can be both warm (friendly, trustworthy) and dominant (assertive, competitive), or exhibit either characteristic independently [48, 56, 20, 18]. This distinction is particularly relevant for AI agents, which can be designed to balance these seemingly contradictory approaches to any arbitrary level, from cold and dominant to warm and dominant, and from cold and submissive to warm and submissive. Our focus on warmth and dominance as organizing dimensions also builds on foundational work in negotiation theory [18, 20, 57], particularly the Dual Concern Model [58, 48, 59]. However, AI-specific capabilities give rise to a vast space of AI-specific negotiation strategies that could drive performance in AI negotiations with no basis in classic interpersonal or negotiation theory. In a tournament setting, unexpected strategies can emerge, succeed, and thus be discovered.

Related work. Unfortunately, this work has yet to incorporate the nearly 70-year history of human negotiations research, including theories about cooperation vs. competition [17, 18], strategic interactions [19], principled negotiation [20], value creation and claiming [21], cognitive biases [22], social perception [23], emotional expression [24], as well as subjective value [25].

Axelrod’s tournament approach led to novel and counterintuitive insights about the evolution of cooperation, including that cooperation can emerge without central authority, that the likelihood of future interaction makes cooperation more attractive, and that “nice” strategies succeed. We tested similar novel and counterintuitive possibilities in the context of AI negotiation. For example, established negotiation theory suggests that warmth is crucial for fostering counterpart subjective value [25], which may contribute to greater objective value [30], although the effects of warmth on objective value are still debated [31, 32]. In human contexts, warmth facilitates trust-building, increases the willingness to share information, and creates psychological safety—all factors that contribute to successful negotiation outcomes [33, 34, 35, 36]. But it is not clear whether the role of warmth, based on human negotiating contexts, applies to AI negotiation.

On the one hand, many believe that it is not important to treat AI agents warmly, as with their human counterparts, because agents do not have feelings the same way humans have feelings [37]. This perspective has led many to overlook the importance of warmth in human-AI interactions as researchers instead turn to technical optimization, rational calculation, strategic positioning, and computational efficiency to optimize AI agents for negotiation [38, 39, 40]. On the other hand, because many AI agents are trained on humangenerated data, they may exhibit human-like responses to warmth and other social cues [41, 42, 43]. Prior work demonstrates that the apparent personality of robotic and AI agents systematically shapes human users’ responses in negotiation and related tasks in ways that are not wholly reducible to their underlying bargaining strategies [44, 45, 12, 46, 47, 8, 7]. Established negotiation theory also emphasizes the importance of dominance and assertiveness in claiming value [48, 49].

Method. 2 Competition Design To investigate these questions in the context of AI negotiation agents, we facilitated 182,812 negotiations between AI agents across multiple scenarios, with different characteristics and objectives, and analyzed how established negotiation principles translate to AI performance. The AI agents were designed by a diverse group of 286 participants, recruited from LinkedIn and negotiation courses worldwide. This group of participants represents over 40 countries and a broad range of experience in negotiation, AI, and computer science, from academics to practitioners (See Fig. 1 and SI Tab. 1 for more comprehensive demographic information). We scored each of the submitted agent designs on how much they emphasized warmth and dominance on a scale of 0 to 100 using GPT-5.2. Following the existing literature, we defined warmth as acting friendly, sympathetic, or sociable, and demonstrating empathy and nonjudgmental understanding of other people’s needs, interests, and positions, and we defined dominance as acting assertive, firm, or forceful, and advocating for one’s own needs, interests, and positions [54, 55] (see SI Sec. 1E and Fig. S19-20 for more details). We validated the GPT measures using independent ratings from the authors on a subset of the prompts. We also tracked agents’ use of AI-specific negotiation tactics, like chain-of-thought reasoning and prompt injection, which could not possibly apply to classic human-human negotiations.

During the competition, human participants wrote detailed instructions (prompts) for AI agents designed to perform well across a diverse set of negotiation scenarios. Drawing on Axelrod’s approach, the competition followed a round-robin design in which each agent negotiated with every agent, including itself, twice in a distributive buyer-seller negotiation (chair price), as well as two integrative landlord-tenant (rental contract) negotiations and recruiter-job candidate (employment contract) negotiations. The distributive negotiations were adapted from Curhan, Eisenkraft, and Elfenbein (2013) [60] and the integrative negotiations were adapted from Neale (1997) [61]. We evaluated agents across five metrics: 1) value claimed (how much value it captured for itself), 2) value created (the total value or “size of the pie” generated jointly through the negotiation), 3) counterpart subjective value (the impression left on the counterpart following the negotiation) [25], 4) efficiency (the number of negotiating turns required to reach agreement or until the negotiation ended without an agreement), and 5) whether a deal was reached. Participants were informed that the first four would determine performance, with efficiency serving as a tiebreaker, but we include deal completion in our analyses because it sheds light on the mechanisms through which warmth and dominance affect other outcomes. We report the aggregate subjective value score in the main text as a holistic measure of the impression left on counterparts [62, 63, 64, 65], but also present facet-level analyses for the four constituent dimensions of subjective value (instrumental, self, process, and relationship) in the Supplementary Information (see SI Sec. 2A and Fig. S27). To incentivize high-quality agent designs, we offered prizes to top performers, including public recognition, access to an online negotiation training program with AI counterparts (“Mastering Negotiation Skills with AI”), and free admission to the Program on Negotiation (PON) AI Summit.

The competition ran from February 1 to 15, 2025. In a preliminary round, participants generated and tested negotiation agents in an interactive “sandbox” environment hosted on iDecisionGames. In this sandbox, participants designed multiple agents and evaluated how they performed against each other in real-time in a distributive negotiation over the sale of a used lamp, a negotiation exercise based on Curhan, Eisenkraft, and Elfenbein’s (2013) case [60] (see SI Sec. 1C.3 and Fig. S17 for more details). The sandbox environment functioned as “in-sample” training, where participants could refine their prompting strategies.

Limitations. While our work reports on the results of the largest international AI negotiation competition ever conducted, it is not without its limitations, which themselves foreshadow new directions in AI negotiation research. First, our competition exclusively analyzed one-shot negotiations rather than repeated interactions. It is well known that negotiation dynamics change dramatically in repeated interactions [81, 62] and that reciprocity and longterm planning play significant roles in repeated games not found in one-shot settings [82]. One promising avenue for future research thus builds on our evidence of the importance of warmth in reaching agreements.

These findings suggest significant implications for repeated interactions and long-term negotiation strategies involving AI. While our current study examines single-encounter negotiations, the balance between warmth and maximizing value in a single deal raises important questions about optimal strategies over time. In repeated human-AI or AI-AI negotiations with memory capabilities, the impact of warmth may be further amplified or potentially recalibrated. Future research should explore how relationship-building through warmth in initial encounters affects subsequent negotiations when AI agents can reference past interactions.