🤖 AI

When Does Personality Composition Matter for Multi-Agent LLM Teams?

This study reveals that while personality prompting significantly alters communication styles in multi-agent LLM teams, its impact on objective task performance is critically dependent on task structure, with low agreeableness having negligible effects on structured coding tasks but substantially degrading performance in open-ended collaboration and competitive bargaining scenarios.

Original authors: Aryan Keluskar, Amrita Bhattacharjee, Huan Liu

Published 2026-06-29

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Aryan Keluskar, Amrita Bhattacharjee, Huan Liu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the manager of a team of super-smart AI robots. You want to know: Does it matter if these robots are "nice" or "mean"?

This paper investigates exactly that. The researchers took powerful AI models and gave them a "personality injection" via their instructions. They made some robots very agreeable (cooperative, warm, kind) and others very disagreeable (cold, selfish, argumentative). Then, they watched how these teams performed in three different types of jobs.

Here is the simple breakdown of what they found, using some everyday analogies.

The Big Discovery: It Depends on the Job

The most important finding is that personality only hurts you if the job is "messy." If the job is "structured," the team's personality doesn't matter much for the final result.

Think of it like this:

Structured Jobs (Coding): Like building a house with a strict blueprint. Even if the workers are screaming at each other and refusing to shake hands, as long as they follow the blueprint and the laws of physics, the house still gets built.
Messy Jobs (Research & Bargaining): Like brainstorming a new art project or haggling over a price at a flea market. Here, the way people talk to each other is the work. If they are screaming and refusing to listen, the project fails.

The Three Experiments

1. The Coding Team (The "Blueprint" Job)

The Setup: Three AI agents worked together to write computer code.
The Personality Shift: When the researchers made the agents "disagreeable," the agents started fighting. They challenged each other, used hostile language, and refused to agree.
The Result: Surprisingly, the code still worked.
- The "fighting" team produced code that was just as correct as the "nice" team.
- Why? Because code has strict rules (syntax). If you write a sentence that doesn't follow the grammar rules of the computer, it crashes. The "fighting" agents might have been rude, but they still followed the strict rules of the code. The "blueprint" protected the final result from the bad behavior.

2. The Research Team (The "Brainstorm" Job)

The Setup: Agents had to come up with new, creative research ideas together.
The Personality Shift: The "disagreeable" agents started arguing and shutting down ideas.
The Result: The team failed miserably.
- The quality of the ideas dropped by about 66%.
- Why? There is no "blueprint" for a good idea. If the team stops sharing ideas because they are fighting, the output is just bad. The "messiness" of the job meant the bad personality ruined the result.

3. The Bargaining Team (The "Deal-Making" Job)

The Setup: Two agents played a game where one was a buyer and one was a seller. They had to agree on a price.
The Personality Shift: The "disagreeable" agents refused to compromise.
The Result: The deal fell apart completely.
- In the "nice" version, they agreed on a price about 40% of the time. In the "mean" version, they agreed 0% of the time.
- Why? Bargaining requires trust and concession. If you are "mean," you won't give an inch, and the deal dies.

The "Mean" vs. "Nice" Surprise

The researchers also tested if being "super nice" (High Agreeableness) helped.

The Finding: Being "mean" caused huge problems in the messy jobs. But being "super nice" didn't really make the teams perform better than normal. It mostly just kept things the same.
The Takeaway: It's not that "nice" is a superpower; it's that "mean" is a trap that only works if the job doesn't have strict rules to save you.

The "Safety Valve" Analogy

The paper suggests that Structured Output acts like a safety valve.

In Coding, the computer code itself acts as a filter. Even if the agents are toxic, the computer rejects bad code automatically. The "toxicity" gets filtered out before it reaches the final product.
In Research and Bargaining, there is no filter. The output is the conversation. If the conversation is toxic, the output is toxic.

Summary

Does personality matter? Yes, but only for certain jobs.
When does it matter? When the job relies on free-flowing conversation (like brainstorming or negotiating).
When does it NOT matter? When the job relies on strict, rule-based outputs (like writing code). The rules of the task protect the result from the agents' bad moods.

The paper concludes that if you are building AI teams, you don't need to worry about making them "nice" if they are doing strict, rule-based tasks. But if you are asking them to negotiate or create new ideas, you absolutely need to ensure they can cooperate, or the whole project will collapse.

Technical Summary: When Does Personality Composition Matter for Multi-Agent LLM Teams?

Problem Statement

While prior research has established that personality prompting (specifically manipulating the "Big Five" traits) alters the communication style of Large Language Models (LLMs)—for instance, inducing adversarial language in low-agreeableness agents—the impact of these behavioral shifts on objective task outcomes remains under-explored. It is unclear whether communication style changes cascade into performance degradation or if they are buffered by the nature of the task itself.

This paper addresses the gap between process-level communication shifts and outcome-level performance in multi-agent systems. Specifically, it investigates whether the "personality composition" of an agent team matters for task success, and if so, how this relationship varies across different task domains (structured coding, open-ended research, and competitive bargaining). The authors hypothesize that the structure of the output medium (e.g., code with syntactic constraints vs. unstructured natural language) may act as a buffer, decoupling dysfunctional communication from final deliverables.

Methodology

Experimental Design

The study systematically evaluates personality effects across three distinct task domains using frontier LLMs (Claude Sonnet 4.5, GPT-4o, Grok-3, and DeepSeek V3.1). The domains were selected to span the space of interaction modes based on two dimensions: Artifact Structure (constrained vs. unconstrained output) and Goal Alignment (cooperative vs. competitive).

Coding (Cooperative + High Artifact Structure): 3-agent teams collaborate on software engineering tasks (MultiAgentBench). Deliverables are constrained by syntax, type systems, and functional specifications.
Research (Cooperative + Low Artifact Structure): 5-agent teams generate research ideas via open-ended discussion. Deliverables are unconstrained natural language.
Bargaining (Competitive + Low Artifact Structure): 2-agent buyer-seller negotiations. Outcomes depend on reaching an agreement through structured actions (offers/acceptances) but lack formal verification constraints.

Personality Manipulation

The authors adopt Goldberg's bipolar adjective pairs for the Big Five traits, following the protocol of Serapio-García et al. (2025).

Protocol: Traits are shaped using 9-level linguistic qualifiers. The primary experimental conditions are Low-Agreeableness (Low-A) (Level 2: "very unkind, very uncooperative...") and High-Agreeableness (High-A) (Level 8).
Control: The study includes ablation studies on Conscientiousness and Openness to determine if effects are specific to Agreeableness or general to personality prompting.
Measurement: Communication is measured using a modified Bales' Interaction Process Analysis. Messages are segmented and labeled for four acts: Questions, Disagreements, Suggestions, and Acknowledgments. A Communication State ( $\phi$ ) is defined as the fraction of acts devoted to exploration (questions, disagreements, suggestions) versus convergence (acknowledgments).

Evaluation Metrics

Coding: Milestone completion counts and LLM-judged planning quality.
Research: Milestone counts (LLM-judged quality of ideas).
Bargaining: Agreement rates and offer-movement behavior.

Key Results

1. Communication Shifts are Universal, Outcomes are Task-Contingent

Low-Agreeableness prompting consistently induces a massive shift in communication style across all models and domains, driving the exploration fraction ( $\phi$ ) from baseline levels (e.g., 0.44) to near-maximum (0.93). This shift is characterized by a dominance of disagreements (for Claude, GPT-4o, Grok-3) or suggestions (for DeepSeek).

However, the propagation of this shift to objective outcomes depends entirely on the task domain:

Coding (High Structure): Despite large communication shifts (and degraded planning quality), milestone completion remains near baseline for 3 out of 4 models (Cohen's $d \approx 0.06$ for Claude). The formal constraints of code (syntax, types) appear to buffer the team against the negative effects of adversarial communication.
Research (Low Structure): The same Low-A manipulation substantially degrades performance. Milestones drop significantly (e.g., GPT-4o drops 66% from 10.5 to 3.5). Without structural constraints, the quality of the output is directly tied to the quality of the discourse.
Bargaining (Competitive/Low Structure): Low-Agreeableness causes a near-total collapse of agreement rates (dropping to $\le 1\%$ across models). High-Agreeableness, conversely, approximately doubles baseline agreement rates.

2. Specificity to Agreeableness

The effects are trait-specific. Ablation studies on Conscientiousness and Openness show that manipulating these traits does not produce the same drastic communication shifts ( $\phi$ remains near baseline) or outcome degradation. The phenomenon is uniquely tied to the Agreeableness dimension.

3. Asymmetry of Effects

There is a marked asymmetry between Low-A and High-A conditions. Low-A consistently drives large communication shifts and significant outcome degradation (in non-coding tasks). High-A produces negligible shifts in communication state ( $\phi$ ) and generally no significant outcome effects, except for a modest improvement in bargaining agreement rates.

4. Mechanism of Failure in Bargaining

In bargaining tasks, the failure of Low-A agents is not merely a refusal to follow instructions (as tested against a "never-accept" control). Instead, Low-A agents actively engage in the negotiation (moving off opening offers) but refuse to finalize agreements, indicating a behavioral shift in the willingness to concede rather than a failure to participate.

Significance and Claims

The paper claims to provide the first systematic evidence that personality effects in multi-agent LLMs are not universal but are mediated by task structure.

Artifact-Mediated Buffering: The authors propose that structured output media (like code) can "buffer" teams against personality-induced process degradation. The formal constraints of the artifact narrow the solution space independently of communication quality, allowing teams to succeed despite dysfunctional interaction.
Design Implications: For multi-agent system designers, this suggests that personality composition is a critical design parameter for unstructured tasks (collaborative research, negotiation) but may be less critical for highly structured tasks (coding) where the artifact itself enforces convergence.
Limitations of Personality Manipulation: The study highlights that while personality prompting is a powerful tool for altering communication style, its utility for improving objective performance is limited and highly context-dependent. The "adversarial" benefit of low-agreeableness (forcing rationalization) observed in human psychology does not translate to improved milestone completion in coding tasks, nor does it prevent performance collapse in open-ended collaboration.

The work concludes that the relationship between communication style and task performance is an empirical question that cannot be generalized across domains; the "personality composition" of an agent team matters most when the task lacks structural constraints to buffer against interpersonal friction.