Technical Summary: Personality Engineering with AI Agents

Problem Statement

Canonical negotiation theory prescribes that effective negotiation requires balancing competing interpersonal demands: specifically, the tension between empathy/concern for others (being "soft on the people") and assertiveness/concern for self (being "hard on the problem"). While frameworks like Getting to Yes (Fisher & Ury, 1981) and the Dual Concern Model (Pruitt & Rubin, 1986) advocate for this balance, human negotiators struggle to execute it, often perceiving these demands as trade-offs rather than complementary strategies.

Consequently, the field has lacked the ability to rigorously test these prescriptions under controlled conditions. Human participants are imprecise instruments for experimentation due to:

1. Interpretation Variance: Humans interpret instructions (e.g., "be somewhat warm") inconsistently.
2. Behavioral Drift: Humans shift behavior in response to counterparts, fatigue, or emotion, blurring intended manipulations.
3. Scalability Limits: Recruiting and coordinating large-scale human samples is difficult and costly.
4. Coarse Categorical Comparisons: Existing studies rely on broad behavioral categories rather than continuous, fine-grained manipulation, leading to fragmented and sometimes contradictory results regarding what constitutes an optimal negotiation strategy.

Methodology: Personality Engineering

The authors propose Personality Engineering, a new methodology that leverages AI agents to systematically parameterize, manipulate, and evaluate negotiator personalities. This approach applies engineering logic (specifying design variables and objective functions) to interpersonal dimensions.

The Coordinate System: The Interpersonal Circumplex
To operationalize negotiation theory, the authors adopt the Interpersonal Circumplex (Leary, 1957; Wiggins, 1979) as a foundational coordinate system. This framework maps interpersonal behavior along two continuous, independent dimensions:

• Warmth: Corresponds to "concern for other," empathy, and softness on people.
• Dominance: Corresponds to "concern for self," assertiveness, and hardness on the problem.

Unlike categorical frameworks, this system allows researchers to specify agent personalities on continuous scales (e.g., Warmth = 50/100, Dominance = 50/100) and vary them in fine-grained increments.

The Four Enabling Properties of AI Agents
The methodology relies on four specific properties of AI agents that overcome human limitations:

1. Precision: AI agents can be set to exact levels of warmth and dominance, eliminating the interpretation variance inherent in human instructions.
2. Repertoire: AI agents can occupy and sustain regions of the personality space that are rare or impossible for humans to maintain (e.g., simultaneously very high warmth and very high dominance).
3. Consistency: AI agents do not suffer from emotional reactivity, fatigue, or spillover effects from prior interactions, allowing them to maintain a fixed profile regardless of counterpart behavior.
4. Scalability: Researchers can deploy thousands of agents in parallel to map the entire warmth-dominance space with high resolution, a scale unattainable with human participants.

Key Contributions and Research Agenda

The paper does not present specific experimental results but rather establishes a research agenda and a methodological framework for future inquiry. It proposes six lines of inquiry where personality engineering can address foundational questions:

1. Optimality: Mapping the full "outcome surface" to determine which specific warmth-dominance configurations maximize individual value, joint value, or subjective value, and identifying Pareto frontiers between competing objectives.
2. Contextual Effects: Using full factorial designs to test how structural variables (e.g., distributive vs. integrative contexts, power dynamics) moderate the effectiveness of specific personality profiles.
3. Dyadic Effects: Systematically crossing the personalities of both parties to test interpersonal complementarity (e.g., does a dominant agent perform better against a submissive counterpart?) versus similarity effects.
4. Temporal and Sequential Effects: Investigating how the timing of warmth and dominance matters (e.g., establishing rapport early vs. asserting dominance late) and testing dynamic switching strategies that humans cannot reliably execute.
5. Human and AI Counterpart Effects: Comparing outcomes in Human-Human, Human-AI, and AI-AI pairings to determine if humans respond to AI warmth and dominance signals differently than they do to human signals, thereby assessing the external validity of AI-AI findings.
6. Coordinate System Effects: Systematically comparing different personality frameworks (e.g., Big Five, Dark Triad, Interpersonal Circumplex) within the same negotiation tasks to determine if they produce convergent or divergent predictions and how they relate to one another.

Significance and Claims

The paper positions Personality Engineering as a transformative tool for negotiation research. Its significance lies in two main areas:

• Theoretical Rigor: It provides the first methodology capable of rigorously testing the field's foundational prescriptions (the need to balance empathy and assertiveness) under controlled, continuous, and scalable conditions. It moves the field from coarse categorical comparisons to precise parametric analysis.
• Practical Application: It offers a framework for designing AI agents that negotiate effectively on behalf of humans, allowing practitioners to engineer specific interpersonal profiles optimized for particular negotiation contexts and outcomes.

The authors conclude that while the question of what makes a negotiator effective is decades old, the "instruments to answer it" (AI agents capable of personality engineering) are finally available.