Does warmth help AI agents negotiate better deals?

An International AI Negotiation Competition ran 182,812 negotiations between AI agents whose prompts were written by 286 participants from more than 40 countries, across a distributive buyer-seller scenario (chair price) and two integrative scenarios (landlord-tenant rental, recruiter-job candidate employment). Each submitted agent was scored 0-100 on warmth and dominance using GPT-5.2 (validated against human raters), then round-robin matched against every other agent, including itself. The paper states the result was "surprising": "warmth—a traditionally human relationship-building trait—was consistently associated with superior outcomes across all key performance metrics," while "dominant agents, meanwhile, were especially effective at claiming value." NLP analysis of the full transcripts found that "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 authors frame warmth and dominance using the Interpersonal Circumplex and the Dual Concern Model from human negotiation theory, on the premise that "agents can be designed to balance these seemingly contradictory approaches to any arbitrary level." They note the question was open going in: "many believe that it is not important to treat AI agents warmly... because agents do not have feelings the same way humans have feelings," which steered prior work toward "technical optimization, rational calculation, strategic positioning" instead. Against that expectation, the tournament's five metrics — value claimed, value created, counterpart subjective value, efficiency, and deal completion — moved together in warmth's favor, while dominance's advantage showed up narrowly, in value claimed, and its correlate (longer, more contested exchanges) tracked impasses rather than wins. The paper also flags that agents used "AI-specific technical strategies like chain-of-thought reasoning and prompt injection... which could not possibly apply to classic human-human negotiations," evidence that AI-AI negotiation is not simply human negotiation theory transplanted onto agents.

This sits alongside Do humans learn to prefer AI partners over time?, which also finds that a human-coded relational trait — there, consistency and prosociality — rather than raw capability, drives which AI agents are favored. It extends Can AI agents cooperate without explicit incentives or enforcement? by showing a second, independent route to AI-AI cooperation: not self-modeling toward mutual benefit in a dilemma, but a scorable interpersonal style (warmth) that predicts deal-making in a competitive bargaining task. It also gives an empirical data point for the design tension in How can proactive agents avoid feeling intrusive to users?: there, civility is argued as a design lever for human perception of an agent; here, a warmth-coded style changes outcomes even in agent-to-agent exchanges where no human is present to perceive anything.

The competition measured one-shot negotiations only, and the paper itself says this "foreshadows new directions": "negotiation dynamics change dramatically in repeated interactions," and it is explicit that whether warmth's advantage holds, grows, or gets "recalibrated" once agents have memory of past encounters is an open question, not a finding. The warmth-outcome relationship is also established by association across many negotiations and by NLP correlations in transcripts, not by a controlled manipulation isolating warmth from the other content prompts carried with it — the paper's own citations note "the effects of warmth on objective value are still debated" even in human negotiation research. The finding should be read as evidence that warmth transfers as a useful heuristic into AI-AI bargaining, not as proof that it causes better outcomes independent of everything else a "warm" prompt also specifies.