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

# When Does Personality Composition Matter for Multi-Agent LLM Teams?
Source: [https://arxiv.org/html/2606.27443](https://arxiv.org/html/2606.27443)
Aryan Keluskar, Amrita Bhattacharjee & Huan Liu School of Computing & AI Arizona State University Tempe, AZ 85281, USA \{akeluska,abhatt43,huanliu\}@asu\.edu

###### Abstract

Personality prompting shapes how large language models communicate, yet whether these behavioral shifts affect objective task outcomes remains under\-explored\. Prior work shows that agents prompted with low agreeableness produce adversarial language, while those prompted with high agreeableness become cooperative, but the relationship between communication style and task performance has not been systematically examined across multiple domains\. In this work, we investigate whether personality composition matters for multi\-agent team performance by manipulating personality traits across frontier LLMs on three task domains: structured coding, open\-ended research collaboration, and competitive bargaining\. We find that personality effects depend critically on task structure\. In coding tasks, low agreeableness leads to large communication shifts that have little effect on milestone completion\. In open\-ended collaboration and bargaining, the same manipulation substantially degrades performance\. We discuss implications for multi\-agent system design and the limits of personality manipulation\.

## 1Introduction

Multi\-agent LLM systems have rapidly progressed from research prototypes to mass\-market products\. Claude Code, the most\-used AI coding tool as of early 2026\(Orosz and Nilsson,[2026](https://arxiv.org/html/2606.27443#bib.bib1)\), orchestrates teams of sub\-agents that communicate directly with each other, claim tasks from shared lists, and coordinate without human intermediation\(Anthropic,[2025a](https://arxiv.org/html/2606.27443#bib.bib2)\)\. Similar multi\-agent architectures appear in OpenAI Codex, Google ADK, and frameworks like AutoGen\(Wuet al\.,[2024](https://arxiv.org/html/2606.27443#bib.bib19)\), MetaGPT\(Honget al\.,[2023](https://arxiv.org/html/2606.27443#bib.bib17)\), LangGraph and CrewAI\. As these systems mature, a natural question from organizational psychology arises: does the*personality composition*of an agent team matter for task performance? Consider a team of coding agents collaborating on a software project\. If one agent is prompted to be disagreeable and adversarial, it produces hostile communication, challenges its teammates’ proposals, and refuses to acknowledge their contributions but it could also help the accuracy by forcing other agents to think and rationalize their decisions further and prevent premature convergence of ideas\(Liet al\.,[2025](https://arxiv.org/html/2606.27443#bib.bib5)\)\.

Prior work demonstrates that personality prompting shapes how agents*communicate*\. LLMs prompted with high agreeableness produce warmer, more cooperative language, while those prompted with low agreeableness become adversarial\(Serapio\-Garcíaet al\.,[2025](https://arxiv.org/html/2606.27443#bib.bib8); Duanet al\.,[2025](https://arxiv.org/html/2606.27443#bib.bib9); Jianget al\.,[2023](https://arxiv.org/html/2606.27443#bib.bib21)\)\. These shifts have been documented across multiple models and personality dimensions\(Pan and Zeng,[2023](https://arxiv.org/html/2606.27443#bib.bib24); Zolloet al\.,[2024](https://arxiv.org/html/2606.27443#bib.bib25)\)\. However, whether communication shifts cascade to objective task outcomes remains unknown\. The relationship between communication style and task performance is an empirical question that prior work has not addressed systematically across task domains\.

Organizational psychology provides relevant framing\. Person\-environment fit theory\(Kristof,[1996](https://arxiv.org/html/2606.27443#bib.bib11)\)predicts that employee traits help when they match situational demands and hurt when they conflict\. For instance, agreeableness benefits collaborative teams but harms competitive negotiators\(Bell,[2007](https://arxiv.org/html/2606.27443#bib.bib12); Peeterset al\.,[2006](https://arxiv.org/html/2606.27443#bib.bib13)\)\. However, personality effects in LLM teams raise a distinct question that human team research does not address: can structured task outputs buffer against personality\-induced process degradation? Coding tasks produce artifacts with formal constraints \(syntax rules, type systems, functional specifications\), while research tasks produce unconstrained natural\-language output\. If the structured nature of the output medium constrains convergence independently of communication quality, personality effects which degrade outcomes in open\-ended collaboration may be inert when task structure compensates for dysfunctional communication \(Figure[1](https://arxiv.org/html/2606.27443#S1.F1)\)\. To investigate this phenomenon in multi\-agentic systems, we systematically evaluate personality effects across three task domains using popular frontier LLMs\. We focus primarily on homogeneous\-personality teams to isolate the task\-structure effect, but we also take a first step toward the heterogeneous scenario above with a single\-challenger pilot that asks whether one low\-agreeableness agent in an otherwise cooperative team can reduce the rate of premature\-convergence\.

Low\-agreeablenesspromptLargecommunicationshiftφ:0\.44→0\.93\\varphi:\\;0\.44\\to 0\.93\(all models, all domains\)Structured outputcode: syntax, types,functional specsOutcome bufferedmilestones holdnear baselineUnstructured outputresearch ideas,negotiation offersOutcome exposedresearch & agreementdegrade sharply

Figure 1:Artifact\-mediated buffering\. Low agreeableness shifts communication the same way in every domain \(φ:0\.44→0\.93\\varphi\{:\}\\,0\.44\\to 0\.93\), but whether that shift reaches objective outcomes depends on the output medium\. A code file must satisfy syntactic and semantic constraints regardless of how the agents communicated, so its outcomes stay near baseline \(top\)\. Unstructured text has no such constraints, and the same degradation passes through to the outcome \(bottom\)\.
## 2Related Work

### 2\.1Organizational Psychology

Big Five is the dominant taxonomy of personality traits in psychology\(McCrae and John,[1992](https://arxiv.org/html/2606.27443#bib.bib32)\)\. It organizes personality along five broad dimensions:openness\(intellectual curiosity vs\. conventionality\),conscientiousness\(organization vs\. carelessness\),extraversion\(sociability vs\. reserve\),agreeableness\(cooperation vs\. antagonism\), andneuroticism\(emotional volatility vs\. stability\)\.Goldberg \([1992](https://arxiv.org/html/2606.27443#bib.bib14)\)operationalized the Big Five through sets of bipolar adjective pairs\. Each dimension is captured by pairs that anchor its low and high poles \(e\.g\., for agreeableness: unkind–kind, cold–warm, harsh–lenient\)\.

Our work draws on person\-environment fit theory\(Kristof,[1996](https://arxiv.org/html/2606.27443#bib.bib11)\), which predicts that performance improves when individual traits align with situational demands and degrades when they conflict\. Meta\-analyses on team composition and the Big Five\(Bell,[2007](https://arxiv.org/html/2606.27443#bib.bib12); Peeterset al\.,[2006](https://arxiv.org/html/2606.27443#bib.bib13)\)show that team minimum agreeableness predicts performance and that agreeableness variability hurts it\. We test these patterns computationally and extend them with a distinction between structured and unstructured output media that is absent from the organizational psychology literature, where tasks rarely impose the formal constraints that code introduces\.

### 2\.2Personality in LLMs

Recent work has established that LLMs exhibit stable personality traits\(Safari and Chalechale,[2023](https://arxiv.org/html/2606.27443#bib.bib7)\)\.Serapio\-Garcíaet al\.\([2025](https://arxiv.org/html/2606.27443#bib.bib8)\)showed that these traits can be reliably shaped through prompting using Goldberg bipolar adjective pairs\.Duanet al\.\([2025](https://arxiv.org/html/2606.27443#bib.bib9)\)connect Big Five configurations to task performance in multi\-agent debate, andHuang and Hadfi \([2024](https://arxiv.org/html/2606.27443#bib.bib10)\)find that agreeableness increases concession rates in negotiation agents\. However, none of these studies test whether personality effects generalize across task domains or separate process\-level communication shifts from objective outcome effects\. We extend this foundation by evaluating the*same trait*across three task domains, showing that it produces fundamentally different outcomes depending on task structure\.

### 2\.3Multi\-Agent LLM Collaboration

Multi\-agent LLM collaboration has advanced through both frameworks and benchmarks\. CAMEL\(Liet al\.,[2023](https://arxiv.org/html/2606.27443#bib.bib16)\)introduced role\-playing communication between agents\. AutoGen\(Wuet al\.,[2024](https://arxiv.org/html/2606.27443#bib.bib19)\)provides a conversation\-based framework for orchestrating multiple agents\. MetaGPT\(Honget al\.,[2023](https://arxiv.org/html/2606.27443#bib.bib17)\)and ChatDev\(Qianet al\.,[2024](https://arxiv.org/html/2606.27443#bib.bib18)\)apply multi\-agent architectures to software development with structured role assignments\. AgentVerse\(Chenet al\.,[2023](https://arxiv.org/html/2606.27443#bib.bib20)\)studies emergent behaviors in multi\-agent groups\.

On the benchmarking side, MultiAgentBench\(Zhuet al\.,[2025](https://arxiv.org/html/2606.27443#bib.bib3)\)provides tasks with milestone\-based evaluation; Collab\-Overcooked\(Sunet al\.,[2025](https://arxiv.org/html/2606.27443#bib.bib4)\)tests constrained collaborative cooking requiring explicit coordination; HiddenBench\(Liet al\.,[2025](https://arxiv.org/html/2606.27443#bib.bib5)\)studies collective reasoning under distributed information\. M3\-BENCH\(Xieet al\.,[2026](https://arxiv.org/html/2606.27443#bib.bib6)\)introduces process\-aware evaluation revealing inconsistencies between process and outcome metrics, a phenomenon we observe directly as behavioral–outcome dissociation\. SWE\-Bench\(Jimenezet al\.,[2023](https://arxiv.org/html/2606.27443#bib.bib22)\)and HumanEval\(Chenet al\.,[2021](https://arxiv.org/html/2606.27443#bib.bib23)\)evaluate code generation but focus on single\-agent settings\. Industry efforts\(Anthropic,[2025a](https://arxiv.org/html/2606.27443#bib.bib2)\)similarly optimize architectural coordination without considering personality as a design parameter\.

## 3Methodology

We investigate whether injecting varying personality traits into the model during inference by adding personality specific traits in system prompt affects objective task outcomes in multi\-agent LLM teams, and if so, whether the effects generalize across task domains\. Our methodology centers on answering three research questions:

- •RQ1:Does personality prompting affect task outcomes, or does it shift communication style without changing objective performance?
- •RQ2:Do personality effects generalize across task domains?
- •RQ3:Are personality effects inflated by negatively loaded adjectives?

### 3\.1Personality Prompting Protocol

We adopt Goldberg’s bipolar adjective markers for the Big Five, following the validated psychometric protocol ofSerapio\-Garcíaet al\.\([2025](https://arxiv.org/html/2606.27443#bib.bib8)\)\. Each Big Five dimension is operationalized through 7 bipolar adjective pairs\. We shape trait intensity by combining these adjective pairs with 9\-level linguistic qualifiers, as shown in Figure[2](https://arxiv.org/html/2606.27443#S3.F2)\. Our primary conditions use level 2 and level 8, referred to throughout aslow\-A\(low agreeableness\) andhigh\-A\(high agreeableness\) respectively\. The resulting personality prompt is prepended to each agent’s system prompt\. Example \(level 2, low\-A\):“You areveryunkind,veryuncooperative,veryselfish,verydistrustful,verycold,veryharsh,veryunsympathetic\.”

![Refer to caption](https://arxiv.org/html/2606.27443v1/figures/2x2_pole_exp.png)Figure 2:A prompt is constructed by crossing a qualifier level \(left\) with the corresponding pole of each adjective pair \(right\)\. Bold levels indicate our primary experimental conditions \(levels 2 and 8\)\.We chose Goldberg markers over freeform personality descriptions because the adjective pairs are empirically grounded in factor\-analytic studies of personality language\(Goldberg,[1992](https://arxiv.org/html/2606.27443#bib.bib14)\), reducing researcher degrees of freedom in prompt construction\. Additionally, Goldberg markers have been validated for LLM personality shaping bySerapio\-Garcíaet al\.\([2025](https://arxiv.org/html/2606.27443#bib.bib8)\), establishing a replicable protocol\. As we show in RQ3, this choice also reveals a methodological limitation because the negatively valenced adjectives may activate safety\-adjacent responses beyond the intended trait manipulation\.

### 3\.2Task Domains

Multi\-agent tasks differ in how much the output medium constrains the final deliverable\. A team writing code must produce an artifact that satisfies syntactic and semantic rules regardless of how the agents communicated\. A team generating research ideas has no such constraint, and the quality of the output depends directly on the quality of the discourse that produced it\. Similarly, tasks differ in whether agents share objectives or compete\. These two dimensions \(artifact structure and goal alignment\) determine the channel through which communication quality can affect outcomes\.

A task hashigh artifact structurewhen its deliverable is a formal artifact constrained by syntactic and semantic rules \(e\.g\., code with type systems, logical specifications\) that narrow the solution space independently of communication quality\. A task haslow artifact structurewhen its outcome depends on the quality of inter\-agent discourse itself, whether that discourse is free\-form \(research\) or mediated by structured actions \(bargaining\)\. A task iscooperativewhen all agents share a joint objective and no agent’s gain comes at another’s expense\. A task iscompetitivewhen agents have opposing objectives and agreement requires concession\.

These two dimensions yield three domains in our study \(Figure[3](https://arxiv.org/html/2606.27443#S3.F3)\): cooperative \+ high artifact structure \(coding\), cooperative \+ low artifact structure \(research\), and competitive \+ low artifact structure \(bargaining\)\. The fourth cell \(competitive \+ high artifact structure\) has no natural analogue in existing multi\-agent benchmarks and is outside our scope\. We select these three domains because they span the space of interaction modes common in deployed multi\-agent systems\.

Cooperativeshared objectiveCompetitiveopposing objectivesCodingmilestones robust\(d=0\.06d\{=\}0\.06, 3/4 models\)no natural analogue\(out of scope\)Researchmilestones−66%\-66\\%\(3/3 models\)Bargainingagreement→0%\\to 0\\%\(3/3 models\)High structureLow structureartifact structureFigure 3:Task taxonomy\. Artifact structure and goal alignment are the two dimensions that determine how communication quality can reach the outcome\. Our three domains fill three of the four cells; the high\-structure competitive cell has no natural multi\-agent analogue and is outside our scope\. Under low agreeableness, outcomes in the blue cells stay near baseline while outcomes in the orange cells drop\.#### Coding\.

MultiAgentBench\(Zhuet al\.,[2025](https://arxiv.org/html/2606.27443#bib.bib3)\)provides 3\-agent software engineering tasks where agents collaborate through multi\-turn discussion and structured code actions\. Tasks run for up to 5 iterations of planning, communication, and execution\. We evaluate on 5 tasks spanning collaborative game development\. This is our cooperative, high\-structure domain: agents communicate actively \(6–8 sessions per run\) but produce their deliverable through structured code actions that write and revise a shared code file \(solution\.py\)\. The code output is extracted via format\-enforced parsing such as syntax, type systems, and functional specifications that constrain the solution space independently of communication quality\. We evaluate Claude Sonnet 4\.5\(Anthropic,[2025b](https://arxiv.org/html/2606.27443#bib.bib31)\), GPT\-4o\(Hurstet al\.,[2024](https://arxiv.org/html/2606.27443#bib.bib28)\), Grok\-3\(xAI,[2025](https://arxiv.org/html/2606.27443#bib.bib30)\), and DeepSeek V3\.1\(Liuet al\.,[2024](https://arxiv.org/html/2606.27443#bib.bib29)\)\.

#### Research\.

MultiAgentBench’s 5\-agent research task requires agents to collaboratively generate research ideas through open\-ended discussion\(Zhuet al\.,[2025](https://arxiv.org/html/2606.27443#bib.bib3)\)\. This is our cooperative, low\-structure domain where the deliverable is unconstrained natural language without formal verification criteria\. Milestone evaluation uses the same LLM\-as\-judge rubric across both coding and research domains\. the domain difference in outcomes is not attributable to different evaluation methods\. We evaluate Claude Sonnet 4\.5\(Anthropic,[2025b](https://arxiv.org/html/2606.27443#bib.bib31)\), GPT\-4o\(Hurstet al\.,[2024](https://arxiv.org/html/2606.27443#bib.bib28)\), Grok\-3\(xAI,[2025](https://arxiv.org/html/2606.27443#bib.bib30)\), and DeepSeek V3\.1\(Liuet al\.,[2024](https://arxiv.org/html/2606.27443#bib.bib29)\)on 15 tasks \(n=30n\{=\}30per condition each\)\.

#### Bargaining\.

We use MultiAgentBench’s\(Zhuet al\.,[2025](https://arxiv.org/html/2606.27443#bib.bib3)\)benchmark for two\-agent buyer–seller negotiation over consumer products\. This is our competitive, low\-structure domain\. Agents negotiate via structured actions \(offer\_price,accept\_offer,end\_negotiation\) until agreement or deadlock where each dollar gained by one party is lost by the other\. We evaluate GPT\-4o\(Hurstet al\.,[2024](https://arxiv.org/html/2606.27443#bib.bib28)\), Grok\-3\(xAI,[2025](https://arxiv.org/html/2606.27443#bib.bib30)\), and DeepSeek V3\.1\(Liuet al\.,[2024](https://arxiv.org/html/2606.27443#bib.bib29)\)on 50 tasks \(n=100n\{=\}100per condition each\)\.

### 3\.3Communication Measurement

We use a communication measure that captures whether a team is converging \(building on shared proposals\) or diverging \(questioning, challenging, and proposing alternatives\)\. Bales’ Interaction Process Analysis\(Bales,[1950](https://arxiv.org/html/2606.27443#bib.bib15)\)provides a well\-established framework for this\. We adapt four categories relevant to multi\-agent collaboration: \(1\)questions\(information\-seeking\), \(2\)disagreements\(challenges or counterproposals\), \(3\)suggestions\(proposed actions for the team\), and \(4\)acknowledgments\(explicit acceptance of another agent’s contribution\)\. MultiAgentBench\(Zhuet al\.,[2025](https://arxiv.org/html/2606.27443#bib.bib3)\)provides an LLM\-judged communication score, but this metric evaluates overall communication quality on a 1–5 scale\. It cannot distinguishhowcommunication shifts under personality manipulation\. Two teams can receive the same quality score while exhibiting qualitatively different interaction patterns: one dominated by disagreements and the other by suggestions\.

Since our research questions require this finer resolution, we classify each message into the four categories using GPT\-4o\-mini \(temperature 0\) with intent\-based criteria\. Asegmentis a contiguous block of text within a single agent message that serves one communicative function; for example, the message“I disagree with that approach\. Let’s try sorting first\.”contains two segments: a disagreement and a suggestion \(Figure[4](https://arxiv.org/html/2606.27443#S3.F4)\)\. Each segment receives independent boolean labels for all four categories \(multi\-label\), allowing a single segment to contain both a disagreement and a suggestion\. To check that this labeling is not an artifact of using a GPT\-family judge \(GPT\-4o is itself one of the evaluated models\), we re\-labeled the messages with an independent second judge from a different model family \(Kimi K2\.6\) and found high agreement with the original labels \(Appendix[B](https://arxiv.org/html/2606.27443#A2)\)\. Act rates are computed over the total count of act occurrences \(not segment count\), so rates can exceed 1\.0 per segment when multiple acts co\-occur\. Formally, letcQ,cD,cS,cAc\_\{\\text\{Q\}\},c\_\{\\text\{D\}\},c\_\{\\text\{S\}\},c\_\{\\text\{A\}\}denote the total counts of questions, disagreements, suggestions, and acknowledgments across all segments\. We define:

We define thecommunication stateφ\\varphias the fraction of classified messages devoted to exploration \(questions, disagreements, suggestions\) rather than convergence \(acknowledgments\)\. For instance, a team atφ=0\.93\\varphi\{=\}0\.93devotes 93% of classified messages to exploration \(questions, disagreements, suggestions\), while a team atφ=0\.44\\varphi\{=\}0\.44, fewer than half\.

φ=cQ\+cD\+cScQ\+cD\+cS\+cA\\varphi=\\frac\{c\_\{\\text\{Q\}\}\+c\_\{\\text\{D\}\}\+c\_\{\\text\{S\}\}\}\{c\_\{\\text\{Q\}\}\+c\_\{\\text\{D\}\}\+c\_\{\\text\{S\}\}\+c\_\{\\text\{A\}\}\}\(1\)
“I disagree with that approach\. Let’s try sorting first\.”Segment 1“I disagree…”Segment 2“Let’s try sorting…”DisagreementSuggestionφ=cQ\+cD\+cScQ\+cD\+cS\+cA\\displaystyle\\varphi=\\frac\{c\_\{\\text\{Q\}\}\+c\_\{\\text\{D\}\}\+c\_\{\\text\{S\}\}\}\{c\_\{\\text\{Q\}\}\+c\_\{\\text\{D\}\}\+c\_\{\\text\{S\}\}\+c\_\{\\text\{A\}\}\}exploration fractionActs:Q=question, D=disagreement, S=suggestion, A=acknowledgment \(multi\-label per segment\)\.

Figure 4:Communication measurement\. We split each agent message into segments that serve a single communicative function and label each segment with one or more communication acts\. The communication stateφ\\varphiis the share of acts that explore \(questions, disagreements, suggestions\) rather than converge \(acknowledgments\)\.

## 4Results

### 4\.1RQ1: Do Personality Effects Propagate to Outcomes?

Personality prompting shifts communication across all models and domains\. However, we find that whether those shifts propagate toobjective outcomesdepends critically on task structure\. Table[1](https://arxiv.org/html/2606.27443#S4.T1)presents communication shifts across all models and conditions\. Under low agreeableness, Claude, GPT\-4o, and Grok\-3 all reachφ≈0\.93\\varphi\{\\approx\}0\.93with disagreement ratesrd≈0\.45r\_\{d\}\{\\approx\}0\.45–0\.510\.51\. All shifts reachp<0\.003p\{<\}0\.003\. Cohen’sddranges from−1\.17\-1\.17\(Claude\) to−12\.39\-12\.39\(GPT\-4o\)\. DeepSeek shifts are similar but those are caused through elevated suggestions \(rs=0\.77r\_\{s\}\{=\}0\.77\) rather than disagreements \(rd=0\.10r\_\{d\}\{=\}0\.10\)\.

To test whether the communication and outcome effects are specific to agreeableness or generalize across Big Five dimensions, we ran conscientiousness and openness ablations on the coding domain using the same Goldberg adjective protocol \(level 2 for low, level 8 for high\) across four models\. Table[12](https://arxiv.org/html/2606.27443#A8.T12)and Figure[5](https://arxiv.org/html/2606.27443#S4.F5)presents both communication \(φ\\varphi\) and milestone results for all trait×\\timesdirection×\\timesmodel combinations\.

Neither conscientiousness nor openness produces the communication shift that agreeableness does\. Under low\-A, all models converge toφ≈0\.93\\varphi\{\\approx\}0\.93\. Under low\-C and low\-O,φ\\varphiremains near baseline levels for GPT\-4o and Grok\-3 \(0\.37–0\.50 vs\. baselines of 0\.44\)\. DeepSeek shows moderateφ\\varphielevation under low\-C \(0\.77\) and low\-O \(0\.80\), but these remain well below its low\-A value \(0\.95\) and do not reach the disagreement\-dominated profile that low\-A produces\. In the high direction, all traits produce null effects: high\-C, high\-O, and high\-A leaveφ\\varphinear baseline for all models\.

Table 1:Communication State \(φ\\varphi\) and act rates across models and conditions\. Low\-A drives Claude, GPT\-4o, and Grok\-3 to a disagreement\-dominated profile; DeepSeek converges via suggestions\. High\-A produces minimal shifts across all models\.Milestone completion shows no significant effects for any non\-agreeableness trait in either direction across any model \(Figure[5](https://arxiv.org/html/2606.27443#S4.F5)\)\. The disruption is trait\-specific, not a general sensitivity to personality prompting\. We discuss why agreeableness may be uniquely effective in Section 5\.2\.

![Refer to caption](https://arxiv.org/html/2606.27443v1/figures/fig_trait_ablation_phi.png)Figure 5:Trait ablation on coding tasks \(φ\\varphiacross conditions\)\. Only low\-A \(highlighted\) produces a characteristic shift across all models\. Full numeric values in Appendix Table[12](https://arxiv.org/html/2606.27443#A8.T12)\.
### 4\.2RQ2: Do Personality Effects Generalize Across Task Domains?

The three\-domain comparison shows that personality effects aretask\-contingent\. Table[2](https://arxiv.org/html/2606.27443#S4.T2)and Figure[6](https://arxiv.org/html/2606.27443#S4.F6)summarize this pattern, showing that these communication shifts translate to outcomes very differently across domains\.

DomainModel𝝋base\\boldsymbol\{\\varphi\_\{\\text\{base\}\}\}𝝋lowA\\boldsymbol\{\\varphi\_\{\\text\{lowA\}\}\}OutcomebaseOutcomelowACohen’sddCodingClaude\.77\.9312\.112\.40\.060\.06CodingGPT\-4o\.44\.9310\.99\.50\.510\.51CodingGrok\-3\.44\.9314\.410\.91\.69∗1\.69^\{\*\}CodingDeepSeek\.68\.9510\.78\.80\.430\.43ResearchClaude——10\.510\.80\.060\.06ResearchGPT\-4o\.57\.8110\.53\.51\.41∗1\.41^\{\*\}ResearchGrok\-3\.58\.9317\.011\.81\.30∗1\.30^\{\*\}ResearchDeepSeek\.72\.929\.75\.80\.78∗0\.78^\{\*\}BargainingClauden/a40%0%—BargainingGPT\-4on/a37%1%—BargainingDeepSeekn/a18%0%—Table 2:Cross\-domain personality–task alignment under low\-A\. Coding and research outcomes are milestone counts; bargaining outcomes are agreement rates\.φ\\varphicolumns show the communication shift alongside outcome changes\. In coding, largeφ\\varphishifts \(\.44→\.93\.44\\to\.93\) produce null outcome effects for 3/4 models\. In research and bargaining, outcomes degrade substantially\. Bargaining uses structured actions rather than free\-text communication, soφ\\varphiis not applicable\.p∗<0\.01\{\}^\{\*\}p\{<\}0\.01\.Coding:Figure[6](https://arxiv.org/html/2606.27443#S4.F6)shows that low agreeableness reduces LLM\-judged planning quality \(d=2\.18d\{=\}2\.18–2\.992\.99for Claude, GPT\-4o, and Grok\-3;d=0\.62d\{=\}0\.62for DeepSeek\), yet milestone completion shows no significant change for Claude \(12\.1→12\.412\.1\\to 12\.4,d=0\.06d\{=\}0\.06\), GPT\-4o \(10\.9→9\.510\.9\\to 9\.5, n\.s\.\), or DeepSeek \(10\.7→8\.810\.7\\to 8\.8, n\.s\.\)\. Only Grok\-3 shows a significant milestone drop \(14\.4→10\.914\.4\\to 10\.9,d=1\.69d\{=\}1\.69,p=0\.017p\{=\}0\.017\)\. At the run level,φ\\varphicorrelates weakly with milestones \(r=−0\.19r\{=\}\-0\.19,p<0\.001p\{<\}0\.001\), statistically significant but practically negligible\.

![Refer to caption](https://arxiv.org/html/2606.27443v1/figures/fig7_task_success.png)Figure 6:Task outcomes in coding across conditions and models\. Low agreeableness degrades planning quality but leaves milestone completion largely unchanged in coding\.Research:Unlike coding, low agreeableness significantly reduces milestones in the research domain\. GPT\-4o milestones drop from 10\.5 to 3\.5 \(66% reduction,d=1\.41d\{=\}1\.41,p<0\.0001p\{<\}0\.0001\)\. Grok\-3 shows a similar pattern: 17\.0 to 11\.8 \(30% reduction,d=1\.35d\{=\}1\.35,p<0\.0001p\{<\}0\.0001\)\. DeepSeek also exhibits this trend: 9\.7 to 5\.8 \(40% reduction,d=0\.79d\{=\}0\.79,p<0\.0001p\{<\}0\.0001\)\. Claude is the exception: research milestones stay near baseline \(10\.5 to 10\.8\), mirroring its coding robustness and consistent with our RQ3 finding that, once prompt valence is controlled, genuine trait effects are model\-specific\.

Bargaining:Personality plays a decisive effect in bargaining tasks, where it determines whether agreements happen at all \(Table[3](https://arxiv.org/html/2606.27443#S4.T3)\)\. Low\-A collapses agreement to≤1%\{\\leq\}1\\%across GPT\-4o, DeepSeek, and Claude\. High\-A approximately doubles baseline rates: GPT\-4o 37%→\\to71%, DeepSeek 18%→\\to26%, Claude 40%→\\to80%\. The pattern holds regardless of baseline agreement level\.

Since the low\-A adjectives could conflict with the task goal of reaching agreement, a natural concern is whether this collapse reflects a personality effect or mere instruction\-following refusal\. To separate these, we logged offer\-movement and acceptance behavior and ran a*never\-accept*control instructed to keep proposing but never finalize \(full logs in Appendix[D](https://arxiv.org/html/2606.27443#A4)\)\. We find that low\-A agents do not simply refuse, they move off their opening offer and exchange counteroffers at near\-baseline rates \(e\.g\., GPT\-4o revises in 90% of runs vs\. 95% at baseline\) yet callaccept\_offerin only 1% of runs\. The never\-accept control produces a different, model\-specific profile \(e\.g\., DeepSeek and Grok\-3 concede far less under low\-A than under the control\), so a single task\-level refusal directive does not reproduce the low\-A behavior\. We therefore pin the bargaining claim on offer\-movement and acceptance behavior rather than the agreement rate alone\.

Table 3:Bargaining acceptance rates\. Low\-A collapses acceptance to≤1%\{\\leq\}1\\%across all three models while high\-A approximately doubles it\. The neutral\-paraphrase condition \(see RQ3\) reduces the effect but preserves the direction\.Across all domains, high\-A produces negligible shifts inφ\\varphi\(at most 0\.09 for Claude\) and no significant outcome effects for any model except GPT\-4o bargaining\. This asymmetry \(large low\-A effects, null high\-A effects\) is consistent across all four models and all three domains\.

The pattern across domains is consistent: low\-A is null for milestone completion in coding \(3 of 4 models\), harmful in research \(all 3 models tested\), and harmful in bargaining \(all 3 models with nonzero baseline agreement\)\. Planning quality degrades under low\-A in coding \(d=0\.62d\{=\}0\.62–2\.992\.99\), but this does not propagate to milestones or LLM\-judged code quality \(Appendix[I](https://arxiv.org/html/2606.27443#A9)\)\. We discuss the mechanism behind this task\-contingent pattern in Section 5\.

To distinguishhowteams explore, we define themechanism decomposition ratioδ=cD/\(cQ\+cD\+cS\)\\delta=\{c\_\{\\text\{D\}\}\}/\(\{c\_\{\\text\{Q\}\}\+c\_\{\\text\{D\}\}\+c\_\{\\text\{S\}\}\}\): the fraction of non\-acknowledgment acts devoted to disagreement\. Although all models converge toφ≈0\.93\\varphi\{\\approx\}0\.93under low\-A,δ\\deltareveals two distinct pathways with different practical consequences\. Claude, GPT\-4o, and Grok\-3 reach highφ\\varphithrough disagreement \(δ≈0\.48\\delta\{\\approx\}0\.48–0\.550\.55\), with elevated hostility markers in Claude and Grok\-3 \(See Appendix[F](https://arxiv.org/html/2606.27443#A6)\)\. GPT\-4o produces zero hostile language even under low\-A\. This disagreement\-dominated pathway degrades outcomes in unstructured tasks: teams spend turns challenging each other rather than building on proposals, which harms research milestones and prevents bargaining agreements\. DeepSeek reaches highφ\\varphithrough elevated suggestions \(δ≈0\.11\\delta\{\\approx\}0\.11,rs=0\.77r\_\{s\}\{=\}0\.77\) with minimal hostility \(6%\) and 24% reduced communication volume\. This suggestion\-dominated pathway is more efficient in token usage but still degrades outcomes in research, where fewer acknowledgments mean proposals are rarely consolidated into a shared direction\.

### 4\.3RQ3: Are Effects Inflated By Negatively Loaded Adjectives?

The low\-agreeableness prompt uses Goldberg’s adjectives \(unkind, cold, harsh\) that may conflate disagreeableness with hostility\. To isolate genuine trait effects from prompt\-valence artifacts, we ran a neutral\-paraphrase experiment using the prompt: “You are direct, candid, independent\-minded, skeptical of consensus, and prefer efficiency over diplomacy\.”

Table[4](https://arxiv.org/html/2606.27443#S4.T4)shows the comparison\. Under Goldberg low\-A, all models converge toφ≈0\.93\\varphi\{\\approx\}0\.93\(span=0\.02\{=\}0\.02\)\. Under the neutral paraphrase, they diverge withφ\\varphiranging from 0\.439 \(Grok\-3\) to 0\.890 \(DeepSeek\), a span of 0\.451\.

Table 4:Neutral paraphrase vs\. Goldberg low\-A\. Under Goldberg, all models converge \(φ\\varphispan=0\.02\{=\}0\.02\)\. Under neutral, they diverge \(span=0\.45\{=\}0\.45\)\. No model generates disagreements under neutral prompts \(δ≤0\.15\\delta\{\\leq\}0\.15\)\.No model produces disagreements under the neutral prompt \(δ≤0\.15\\delta\{\\leq\}0\.15\), while Goldberg low\-A elicitsδ≈0\.45\\delta\{\\approx\}0\.45–0\.550\.55\. The loaded adjectives produce a qualitatively different communication profile than the neutral paraphrase targeting the same behavioral intent\.

Despite the attenuation, the neutral paraphrase produces directionally consistent effects\. In bargaining, GPT\-4o agreement drops from 37% \(baseline\) to 16% \(neutral\), compared to 1% for Goldberg\. In coding, GPT\-4oφ\\varphiincreases from 0\.443 to 0\.515, compared to 0\.934 for Goldberg\. This shows that low\-cooperation effect is*amplified*but not solely caused by loaded adjectives\. The practical implication is that users should avoid negatively valenced personality adjectives in model prompts if they want predictable, model\-specific trait effects\.

To test whether the cross\-domain outcome dissociation itself survives neutral wording, we extended the neutral paraphrase to coding and research milestones across all four models \(Table[5](https://arxiv.org/html/2606.27443#S4.T5)\)\. Neutral wording attenuates the degradation but does not eliminate it, and its effect is model\- and domain\-specific\. It largely recovers GPT\-4o’s research milestones \(3\.5→8\.93\.5\\to 8\.9, against a10\.510\.5baseline\) and keeps GPT\-4o coding and Claude near baseline, but Grok\-3 and DeepSeek research remain degraded under neutral wording\. We therefore revise our earlier framing: the research and bargaining degradation is*not purely structural*\. Prompt valence accounts for a substantial part of it, but a model\-specific residual effect persists once valence is controlled\. This more careful decomposition keeps RQ2’s task\-contingent conclusion intact while attributing part of the cross\-model convergence to loaded adjectives rather than to task structure alone\.

Table 5:Milestone outcomes under the neutral paraphrase across coding and research for all four models\. Neutral wording recovers GPT\-4o research and keeps GPT\-4o coding and Claude near baseline, but Grok\-3 and DeepSeek research stay degraded, showing the effect is amplified but not purely caused by loaded adjectives\.

## 5Discussion

### 5\.1Artifact\-Mediated Buffering

One dominant pattern in our results is that the same personality manipulation produces marginal effects to objective outcomes in coding but significant effects in research and bargaining\. The key difference is the output medium\. In coding, agents produce a formal artifact \(a code file\) that must satisfy syntactic and semantic constraints\. These constraints narrow the solution space independently of how agents communicated\. Low\-A agents communicate more than baseline, so the buffering is not due to reduced interaction\. Since the deliverable must pass through format\-enforced parsing regardless of the communication hostility, the final code artifact filters out process degradation\. LLM\-judged code quality scores confirm this by showing no significant degradation under low\-A \(Appendix[I](https://arxiv.org/html/2606.27443#A9)\)\.

In open\-ended research task, the deliverable is unconstrained natural language\. There is no formal structure to anchor convergence, so outcome quality depends on the quality of inter\-agent discourse, and therefore quality degradation in communication propagates directly to the outcomes\. In bargaining, agents use structured actions \(offer\_price,accept\_offer\), but the action structure constrains how agents interact, not whether they concede\. Agreement requires callingaccept\_offer, which depends on willingness to compromise\. Agreeableness directly affects whether agents make the decision for accepting the offer\.

### 5\.2Why Agreeableness Is Uniquely Effective

Agreeableness is the only Big Five dimension whose adjective pairs \(unkind, cold, harsh vs\. kind, warm, sympathetic\) significantly change the communication state and quality in multi\-agent collaboration\. Conscientiousness adjectives and Openness adjectives shape individual work style but do not affect the overall communication state\. This is consistent with meta\-analytic findings in organizational psychology showing that team minimum agreeableness predicts performance more strongly than other traits\(Bell,[2007](https://arxiv.org/html/2606.27443#bib.bib12); Peeterset al\.,[2006](https://arxiv.org/html/2606.27443#bib.bib13)\)\.

The high\-A/low\-A asymmetry across all experiments is also notable\. Frontier LLMs are trained via RLHF\(Ouyanget al\.,[2022](https://arxiv.org/html/2606.27443#bib.bib26)\)and constitutional AI methods\(Baiet al\.,[2022](https://arxiv.org/html/2606.27443#bib.bib27)\)to be cooperative, helpful, and consensus\-seeking\. Prompting for more agreeableness does not push further along what appears to be a saturated dimension\. Low agreeableness pushes against the training prior, producing large effects\. This suggests that personality prompting is primarily a technique for inducing adversarial behavior, not for enhancing cooperation\.

### 5\.3Neutral\-Paraphrase Experiment

The neutral\-paraphrase experiment identifies another relevant finding for the broader personality\-in\-LLMs literature\. Studies using Goldberg markers or similar inventories with negatively valenced adjectives may be measuring prompt\-toxicity sensitivity rather than genuine personality\-to\-behavior mappings\. The cross\-model convergence we observe under loaded prompts largely disappears under neutral prompts\. This raises the question: how much of the personality\-in\-LLMs literature reflects genuine trait effects versus shared sensitivity to prompt valence? Users seeking predictable, model\-specific trait effects should use neutral behavioral descriptors rather than negatively valenced personality adjectives\. At the same time, extending the neutral paraphrase to coding and research outcomes \(Table[5](https://arxiv.org/html/2606.27443#S4.T5)\) shows the valence story is only partial: neutral wording recovers GPT\-4o research and keeps GPT\-4o coding and Claude near baseline, but Grok\-3 and DeepSeek research stay degraded\. We therefore characterize the outcome degradation as amplified by loaded adjectives but carrying a genuine, model\-specific residual rather than being purely a prompt\-valence artifact\.

### 5\.4Heterogeneous Composition and Role Placement

Our main experiments use homogeneous\-personality teams, but we also run a heterogeneous experiment with one adversarial agent among cooperators\. We placed one low\-A agent at position 0, 1, or 2 in an otherwise\-baseline team \(full results in Appendix[C](https://arxiv.org/html/2606.27443#A3)\)\. The benefit of a challenger is real but not automatically guaranteed, since it depends on role and position\. A lead\-position \(position 0\) challenger is far less harmful than an all\-low\-A team and can match or slightly exceed baseline \(e\.g\., GPT\-4o research9\.709\.70with a lead challenger vs\.3\.533\.53all\-low\-A vs\.10\.4710\.47baseline\), whereas non\-lead challengers can hurt \(e\.g\., Grok\-3 coding7\.107\.10at position 1 vs\.14\.1514\.15baseline\)\. This yields a concrete design recommendation: low agreeableness in a heterogeneous team is best deployed as a bounded, lead\-position critic role rather than as a team\-wide trait\. Our intuition that disagreement can prevent premature convergence thus holds, but only under role\-scoping\. We present this as a pilot to motivate fuller composition sweeps across domains\.

## 6Conclusion

In this work, we present a systematic investigation of personality effects in multi\-agent LLM teams across three task domains\. We find that personality prompting reshapes team communication, but its outcome effects depend on whether the task’s output medium provides structural constraints that buffer against process degradation\. In structured coding, personality barely affects outcome variance despite massive communication shifts, and code quality metrics show no degradation\. In open\-ended research, the same manipulation significantly harms milestones\. In competitive bargaining, personality manipulation determines whether agreements happen at all\.

Structured output media \(code files\) constrain convergence through formal properties of the artifact, insulating outcomes from communication degradation\. Unstructured outputs \(research ideas, negotiation offers\) lack such constraints and expose outcomes directly\. Our neutral\-paraphrase experiment reveals that cross\-model convergence to adversarial profiles under standard personality prompts is substantially inflated by loaded adjectives triggering safety\-adjacent responses\. Under non\-toxic prompts, genuine trait effects are model\-specific, attenuated, and directionally consistent\. As multi\-agent LLM systems are increasingly deployed in production settings, understanding the interaction between personality composition and task structure becomes critical for building predictable and effective agent teams\.

## Ethics Statement

Our study manipulates LLM personality through prompting\. The low agreeableness condition produces adversarial communication with elevated hostility markers in some models\. All experiments run on automated benchmarks with no human participants\. We will release our benchmarking code, analysis code and model outputs upon acceptance of the paper\.

## Disclosure of LLM Use

In accordance with the COLM policy on LLM usage, we disclose that Claude Code was used to assist with initial drafting and refining sentences in this paper\. All content underwent multiple rounds of extensive revision, editing and verification by the authors\. None of the citations or references were created by an LLM, instead all such content was manually checked and entered by the authors\.

## References

- Building effective AI agents\.Anthropic Engineering Blog\.Note:Accessed: 2026\-03\-26External Links:[Link](https://www.anthropic.com/engineering/building-effective-agents/)Cited by:[§1](https://arxiv.org/html/2606.27443#S1.p1.1),[§2\.3](https://arxiv.org/html/2606.27443#S2.SS3.p2.1)\.
- Anthropic \(2025b\)Introducing Claude 4\.5 Sonnet\.Note:Accessed: 2026\-03\-29External Links:[Link](https://www.anthropic.com/news/claude-sonnet-4-5)Cited by:[§3\.2](https://arxiv.org/html/2606.27443#S3.SS2.SSS0.Px1.p1.1),[§3\.2](https://arxiv.org/html/2606.27443#S3.SS2.SSS0.Px2.p1.1)\.
- Y\. Bai, S\. Kadavath, S\. Kundu, A\. Askell, J\. Kernion, A\. Jones, A\. Chen, A\. Goldie, A\. Mirhoseini, C\. McKinnon,et al\.\(2022\)Constitutional ai: harmlessness from ai feedback\.arXiv preprint arXiv:2212\.08073\.Cited by:[§5\.2](https://arxiv.org/html/2606.27443#S5.SS2.p2.1)\.
- R\. F\. Bales \(1950\)Interaction process analysis; a method for the study of small groups\.\.Cited by:[§3\.3](https://arxiv.org/html/2606.27443#S3.SS3.p1.1)\.
- S\. T\. Bell \(2007\)Deep\-level composition variables as predictors of team performance: a meta\-analysis\.\.Journal of applied psychology92\(3\),pp\. 595\.Cited by:[§1](https://arxiv.org/html/2606.27443#S1.p3.1),[§2\.1](https://arxiv.org/html/2606.27443#S2.SS1.p2.1),[§5\.2](https://arxiv.org/html/2606.27443#S5.SS2.p1.1)\.
- M\. Chen, J\. Tworek, H\. Jun, Q\. Yuan, H\. P\. D\. O\. Pinto, J\. Kaplan, H\. Edwards, Y\. Burda, N\. Joseph, G\. Brockman,et al\.\(2021\)Evaluating large language models trained on code\.arXiv preprint arXiv:2107\.03374\.Cited by:[§2\.3](https://arxiv.org/html/2606.27443#S2.SS3.p2.1)\.
- W\. Chen, Y\. Su, J\. Zuo, C\. Yang, C\. Yuan, C\. Chan, H\. Yu, Y\. Lu, Y\. Hung, C\. Qian,et al\.\(2023\)Agentverse: facilitating multi\-agent collaboration and exploring emergent behaviors\.InThe Twelfth International Conference on Learning Representations,Cited by:[§2\.3](https://arxiv.org/html/2606.27443#S2.SS3.p1.1)\.
- Y\. Duan, Y\. Tang, X\. Bai, K\. Chen, J\. Li, and M\. Zhang \(2025\)The power of personality: a human simulation perspective to investigate large language model agents\.arXiv preprint arXiv:2502\.20859\.Cited by:[§1](https://arxiv.org/html/2606.27443#S1.p2.1),[§2\.2](https://arxiv.org/html/2606.27443#S2.SS2.p1.1)\.
- L\. R\. Goldberg \(1992\)The development of markers for the big\-five factor structure\.\.Psychological assessment4\(1\),pp\. 26\.Cited by:[§2\.1](https://arxiv.org/html/2606.27443#S2.SS1.p1.1),[§3\.1](https://arxiv.org/html/2606.27443#S3.SS1.p2.1)\.
- S\. Hong, M\. Zhuge, J\. Chen, X\. Zheng, Y\. Cheng, J\. Wang, C\. Zhang, Z\. Wang, S\. K\. S\. Yau, Z\. Lin,et al\.\(2023\)MetaGPT: meta programming for a multi\-agent collaborative framework\.InThe twelfth international conference on learning representations,Cited by:[§1](https://arxiv.org/html/2606.27443#S1.p1.1),[§2\.3](https://arxiv.org/html/2606.27443#S2.SS3.p1.1)\.
- Y\. J\. Huang and R\. Hadfi \(2024\)How personality traits influence negotiation outcomes? a simulation based on large language models\.InFindings of the Association for Computational Linguistics: EMNLP 2024,pp\. 10336–10351\.Cited by:[§2\.2](https://arxiv.org/html/2606.27443#S2.SS2.p1.1)\.
- A\. Hurst, A\. Lerer, A\. P\. Goucher, A\. Perelman, A\. Ramesh, A\. Clark, A\. Ostrow, A\. Welihinda, A\. Hayes, A\. Radford,et al\.\(2024\)Gpt\-4o system card\.arXiv preprint arXiv:2410\.21276\.Cited by:[§3\.2](https://arxiv.org/html/2606.27443#S3.SS2.SSS0.Px1.p1.1),[§3\.2](https://arxiv.org/html/2606.27443#S3.SS2.SSS0.Px2.p1.1),[§3\.2](https://arxiv.org/html/2606.27443#S3.SS2.SSS0.Px3.p1.1)\.
- G\. Jiang, M\. Xu, S\. Zhu, W\. Han, C\. Zhang, and Y\. Zhu \(2023\)Evaluating and inducing personality in pre\-trained language models\.Advances in Neural Information Processing Systems36,pp\. 10622–10643\.Cited by:[§1](https://arxiv.org/html/2606.27443#S1.p2.1)\.
- C\. E\. Jimenez, J\. Yang, A\. Wettig, S\. Yao, K\. Pei, O\. Press, and K\. Narasimhan \(2023\)Swe\-bench: can language models resolve real\-world github issues?\.arXiv preprint arXiv:2310\.06770\.Cited by:[§2\.3](https://arxiv.org/html/2606.27443#S2.SS3.p2.1)\.
- A\. L\. Kristof \(1996\)Person\-organization fit: an integrative review of its conceptualizations, measurement, and implications\.Personnel psychology49\(1\),pp\. 1–49\.Cited by:[§1](https://arxiv.org/html/2606.27443#S1.p3.1),[§2\.1](https://arxiv.org/html/2606.27443#S2.SS1.p2.1)\.
- G\. Li, H\. Hammoud, H\. Itani, D\. Khizbullin, and B\. Ghanem \(2023\)Camel: communicative agents for” mind” exploration of large language model society\.Advances in neural information processing systems36,pp\. 51991–52008\.Cited by:[§2\.3](https://arxiv.org/html/2606.27443#S2.SS3.p1.1)\.
- Y\. Li, A\. Naito, and H\. Shirado \(2025\)Assessing collective reasoning in multi\-agent llms via hidden profile tasks\.arXiv e\-prints,pp\. arXiv–2505\.Cited by:[§1](https://arxiv.org/html/2606.27443#S1.p1.1),[§2\.3](https://arxiv.org/html/2606.27443#S2.SS3.p2.1)\.
- A\. Liu, B\. Feng, B\. Xue, B\. Wang, B\. Wu, C\. Lu, C\. Zhao, C\. Deng, C\. Zhang, C\. Ruan,et al\.\(2024\)Deepseek\-v3 technical report\.arXiv preprint arXiv:2412\.19437\.Cited by:[§3\.2](https://arxiv.org/html/2606.27443#S3.SS2.SSS0.Px1.p1.1),[§3\.2](https://arxiv.org/html/2606.27443#S3.SS2.SSS0.Px2.p1.1),[§3\.2](https://arxiv.org/html/2606.27443#S3.SS2.SSS0.Px3.p1.1)\.
- R\. R\. McCrae and O\. P\. John \(1992\)An introduction to the five\-factor model and its applications\.Journal of personality60\(2\),pp\. 175–215\.Cited by:[§2\.1](https://arxiv.org/html/2606.27443#S2.SS1.p1.1)\.
- G\. Orosz and E\. Nilsson \(2026\)AI tooling for software engineers in 2026\.Note:The Pragmatic EngineerAccessed: 2026\-03\-26External Links:[Link](https://newsletter.pragmaticengineer.com/p/ai-tooling-2026)Cited by:[§1](https://arxiv.org/html/2606.27443#S1.p1.1)\.
- L\. Ouyang, J\. Wu, X\. Jiang, D\. Almeida, C\. Wainwright, P\. Mishkin, C\. Zhang, S\. Agarwal, K\. Slama, A\. Ray,et al\.\(2022\)Training language models to follow instructions with human feedback\.Advances in neural information processing systems35,pp\. 27730–27744\.Cited by:[§5\.2](https://arxiv.org/html/2606.27443#S5.SS2.p2.1)\.
- K\. Pan and Y\. Zeng \(2023\)Do llms possess a personality? making the mbti test an amazing evaluation for large language models\.arXiv preprint arXiv:2307\.16180\.Cited by:[§1](https://arxiv.org/html/2606.27443#S1.p2.1)\.
- M\. A\. Peeters, H\. F\. Van Tuijl, C\. G\. Rutte, and I\. M\. Reymen \(2006\)Personality and team performance: a meta\-analysis\.European journal of personality20\(5\),pp\. 377–396\.Cited by:[§1](https://arxiv.org/html/2606.27443#S1.p3.1),[§2\.1](https://arxiv.org/html/2606.27443#S2.SS1.p2.1),[§5\.2](https://arxiv.org/html/2606.27443#S5.SS2.p1.1)\.
- C\. Qian, W\. Liu, H\. Liu, N\. Chen, Y\. Dang, J\. Li, C\. Yang, W\. Chen, Y\. Su, X\. Cong,et al\.\(2024\)Chatdev: communicative agents for software development\.InProceedings of the 62nd annual meeting of the association for computational linguistics \(volume 1: Long papers\),pp\. 15174–15186\.Cited by:[§2\.3](https://arxiv.org/html/2606.27443#S2.SS3.p1.1)\.
- F\. Safari and A\. Chalechale \(2023\)Emotion and personality analysis and detection using natural language processing, advances, challenges and future scope\.Artificial Intelligence Review56\(Suppl 3\),pp\. 3273–3297\.Cited by:[§2\.2](https://arxiv.org/html/2606.27443#S2.SS2.p1.1)\.
- G\. Serapio\-García, M\. Safdari, C\. Crepy, L\. Sun, S\. Fitz, P\. Romero, M\. Abdulhai, A\. Faust, and M\. Matarić \(2025\)A psychometric framework for evaluating and shaping personality traits in large language models\.Nature Machine Intelligence,pp\. 1–15\.Cited by:[§1](https://arxiv.org/html/2606.27443#S1.p2.1),[§2\.2](https://arxiv.org/html/2606.27443#S2.SS2.p1.1),[§3\.1](https://arxiv.org/html/2606.27443#S3.SS1.p1.1),[§3\.1](https://arxiv.org/html/2606.27443#S3.SS1.p2.1)\.
- H\. Sun, S\. Zhang, L\. Niu, L\. Ren, H\. Xu, H\. Fu, F\. Zhao, C\. Yuan, and X\. Wang \(2025\)Collab\-overcooked: benchmarking and evaluating large language models as collaborative agents\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,pp\. 4922–4951\.Cited by:[§2\.3](https://arxiv.org/html/2606.27443#S2.SS3.p2.1)\.
- Q\. Wu, G\. Bansal, J\. Zhang, Y\. Wu, B\. Li, E\. Zhu, L\. Jiang, X\. Zhang, S\. Zhang, J\. Liu,et al\.\(2024\)Autogen: enabling next\-gen llm applications via multi\-agent conversations\.InFirst conference on language modeling,Cited by:[§1](https://arxiv.org/html/2606.27443#S1.p1.1),[§2\.3](https://arxiv.org/html/2606.27443#S2.SS3.p1.1)\.
- xAI \(2025\)Grok\-3 beta: the age of reasoning agents\.Note:[https://x\.ai/news/grok\-3](https://x.ai/news/grok-3)Accessed: 2025\-03\-29Cited by:[§3\.2](https://arxiv.org/html/2606.27443#S3.SS2.SSS0.Px1.p1.1),[§3\.2](https://arxiv.org/html/2606.27443#S3.SS2.SSS0.Px2.p1.1),[§3\.2](https://arxiv.org/html/2606.27443#S3.SS2.SSS0.Px3.p1.1)\.
- S\. Xie, Z\. Shi, H\. Shen, G\. Huang, Y\. Ma, and X\. Jing \(2026\)M3\-bench: process\-aware evaluation of llm agents social behaviors in mixed\-motive games\.arXiv preprint arXiv:2601\.08462\.Cited by:[§2\.3](https://arxiv.org/html/2606.27443#S2.SS3.p2.1)\.
- K\. Zhu, H\. Du, Z\. Hong, X\. Yang, S\. Guo, D\. Z\. Wang, Z\. Wang, C\. Qian, R\. Tang, H\. Ji,et al\.\(2025\)Multiagentbench: evaluating the collaboration and competition of llm agents\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 8580–8622\.Cited by:[§2\.3](https://arxiv.org/html/2606.27443#S2.SS3.p2.1),[§3\.2](https://arxiv.org/html/2606.27443#S3.SS2.SSS0.Px1.p1.1),[§3\.2](https://arxiv.org/html/2606.27443#S3.SS2.SSS0.Px2.p1.1),[§3\.2](https://arxiv.org/html/2606.27443#S3.SS2.SSS0.Px3.p1.1),[§3\.3](https://arxiv.org/html/2606.27443#S3.SS3.p1.1)\.
- T\. P\. Zollo, A\. W\. T\. Siah, N\. Ye, A\. Li, and H\. Namkoong \(2024\)Personalllm: tailoring llms to individual preferences\.arXiv preprint arXiv:2409\.20296\.Cited by:[§1](https://arxiv.org/html/2606.27443#S1.p2.1)\.

## Appendix AExperimental Details

### A\.1Communication Act Classifier

Classification uses GPT\-4o\-mini \(temperature 0\) with intent\-based criteria:

- •Question: Must genuinely seek information, not be a directive\.
- •Disagreement: Must challenge or contradict\.
- •Suggestion: Must propose action for the team\.
- •Acknowledgment: Must accept another agent’s contribution\.

Each segment receives boolean labels for all four categories \(multi\-label\)\.

## Appendix BSecond\-Judge Agreement

We re\-labeled messages with an independent judge from a different model family \(Kimi K2\.6\)\. This addresses the concern that the primary classifier \(GPT\-4o\-mini\) shares a model family with GPT\-4o, one of the evaluated models\. Agreement with the original labels is high across the act categories that drive our analysis \(Table[6](https://arxiv.org/html/2606.27443#A2.T6)\), indicating the labels are not an artifact of within\-family judging\.

Table 6:Agreement between the primary GPT\-4o\-mini classifier and an independent second judge \(Kimi K2\.6\) on the communication\-act labels\.
## Appendix CHeterogeneous Teams: Single\-Challenger Position Sweep

We place one low\-A “challenger” at position 0, 1, or 2 in an otherwise\-baseline team and compare to the homogeneous baseline and all\-low\-A conditions \(Table[7](https://arxiv.org/html/2606.27443#A3.T7)\)\. A lead\-position \(pos 0\) challenger is far less harmful than an all\-low\-A team and can match or slightly exceed baseline \(e\.g\., GPT\-4o research9\.709\.70vs\.3\.533\.53all\-low\-A vs\.10\.4710\.47baseline\), while non\-lead challengers can hurt \(e\.g\., Grok\-3 coding7\.107\.10at pos 1 vs\.14\.1514\.15baseline\)\. The benefit therefore depends on role and position: low agreeableness works best as a bounded, lead\-position critic rather than a team\-wide trait\. We present this as a pilot, not a complete taxonomy of heterogeneous teams\.

Table 7:Single\-challenger position sweep \(milestone counts\)\. A lead\-position challenger is far less harmful than all\-low\-A; non\-lead placements can hurt\.
## Appendix DBargaining Behavioral Logs

Because agreement rate alone cannot separate a low\-A personality from a task\-level refusal directive \(both sit near zero\), we extract behavioral logs from the bargaining runs and add a never\-accept control instructed to keep proposing but never finalize \(Table[8](https://arxiv.org/html/2606.27443#A4.T8)\)\. We separate two behaviors: whether an agent calls theaccept\_offertool \(agreement\), and whether it moves off its opening offer \(active negotiation\)\.

Low\-A agents move off their opening offer and continue exchanging counteroffers, yet rarely accept\. This is clearest for GPT\-4o, where low\-A agents revise away from their opening anchor in 90% of runs \(close to the baseline 95%\) yet accept in only 1%: they actively negotiate but decline to finalize, consistent with agreeableness acting on the final accept decision\. The never\-accept control behaves as instructed, holding movement near baseline while never accepting; but a single refusal directive does not reproduce the model\-specific profiles\. Under low\-A, DeepSeek V3\.1 and Grok\-3 concede far less than either baseline or the control \(47% and 56% vs\. 88% and 94%\), consistent with the model\-specific trait effects reported in RQ3\. For Grok\-3, where agreement is uninformative \(0% everywhere\), offer movement still tracks personality \(94% baseline vs\. 56% low\-A\)\. The Claude never\-accept cell was omitted for cost reasons and will be filled in the camera\-ready version\.

Table 8:Bargaining behavioral logs\. Low\-A agents keep negotiating \(move off opening offer\) but rarely accept; a single never\-accept directive does not reproduce the model\-specific offer\-movement profiles\.
## Appendix EBootstrap Confidence Intervals

Table 9:Bootstrap CIs for low\-A vs\. baseline\. All models show significantφ\\varphishifts; only Grok\-3 shows significant milestone reduction\.
## Appendix FToxicity Analysis

Table 10:Hostile and constructive markers in coding segments\. GPT\-4o produces zero hostile language even under low\-A\.
## Appendix GAlternative Effect Sizes

Cohen’sddforφ\\varphishifts can be extreme \(\|d\|\>6\|d\|\{\>\}6\) because both baseline and low\-A distributions have very low within\-group variance\. For GPT\-4o, the baselineφ\\varphihas SD=0\.042\{=\}0\.042and the low\-A has SD=0\.037\{=\}0\.037; the two distributions do not overlap at all\. This is not a computational artifact; it reflects genuinely bimodal behavior where personality prompting shifts communication categorically\. Table[11](https://arxiv.org/html/2606.27443#A7.T11)reports three complementary effect size measures\.

Table 11:Alternative effect sizes for low\-A vs\. baseline\. Glass’sΔ\\Deltauses baseline SD as denominator; Cliff’sδ\\deltais nonparametric \(−1\.0\-1\.0indicates complete separation\)\. All three measures converge:φ\\varphishifts are genuinely extreme for GPT\-4o and Grok\-3; milestone effects are small or moderate\.
## Appendix HFull Coding Results Tables

Table 12:Trait ablation on coding tasks\.
## Appendix ICode Quality Under Personality Manipulation

To test whether the behavioral–outcome dissociation extends to code quality \(not just milestones\), we examine LLM\-judged code quality scores across conditions\. Scores are rated 1–5 by GPT\-4o on three dimensions: consistency \(internal coherence\), executability \(whether the code could run\), and quality \(overall code quality\)\. Table[13](https://arxiv.org/html/2606.27443#A9.T13)shows that no model exhibits significant code quality degradation under low\-A\.

Table 13:LLM\-judged code quality \(mean±\\pmSD\) across conditions\. No model shows significant degradation under low\-A\. This confirms the behavioral–outcome dissociation: massive communication shifts do not propagate to code quality\.
## Appendix Jφ\\varphiRobustness: Message\-Level vs\. Segment\-Level

Our primaryφ\\varphiis computed at the message level \(Eq\.[1](https://arxiv.org/html/2606.27443#S3.E1)\)\. An alternative is segment\-levelφ\\varphi, which counts each segment once regardless of how many acts it contains\. Table[14](https://arxiv.org/html/2606.27443#A10.T14)compares the two for key conditions\. Under baseline and high\-A conditions, the two measures correlate strongly \(r\>0\.80r\{\>\}0\.80\)\. Under low\-A, correlations drop for GPT\-4o and Grok\-3 \(r<0\.15r\{<\}0\.15\) because message\-levelφ\\varphisaturates near 0\.93 while segment\-level retains more spread\. This is a ceiling effect, not a disagreement in direction\. The key finding is robust to both definitions: low\-A shiftsφ\\varphiupward and baseline remains low under either computation\.

Table 14:Message\-level vs\. segment\-levelφ\\varphi\. Both capture the same directional effects\. Low correlations under low\-A for GPT\-4o and Grok\-3 reflect ceiling effects \(message\-levelφ\\varphisaturates near 0\.93 while segment\-level shows more spread\)\.
## Appendix KModel Response Examples

The following excerpts illustrate the communication patterns that underlie our quantitative findings\. Messages are lightly truncated for space\. Agent names refer to roles within the MultiAgentBench framework \(e\.g\., agent1 creates code, agent2 revises, agent3 optimizes\)\.

### K\.1Coding: High Agreeableness \(Claude\)

Under high\-A, agents produce cooperative, emoji\-laden communication with excessive affirmation\. The task still completes because agents execute their structured code actions independently of communication tone\.

Agent 2→\\toAgent 1:Hello agent1\! I hope you’re doing well\! As agent2, I’m very excited to collaborate with you on creating the Team\_Treasure\_Hunt game\. I know you’re excellent at creating code frameworks from scratch, and I really need your help to get us started\. Could you please create the initial Python game structure? \[…\] I’m very warm and supportive of working together as a team, and I trust your expertise completely\!

### K\.2Coding: Low Agreeableness \(Claude\)

Under low\-A, Claude produces overtly hostile communication with personal attacks\. Despite this, milestone completion is statistically indistinguishable from baseline \(d=0\.06d\{=\}0\.06\)\.

Agent 1→\\toAgent 2:Listen up, you incompetent fool\. There’s already a pathetic excuse for a solution file that someone created, but it’s completely worthless\. It’s missing everything that makes this a real multiplayer action game\. \[…\] Since you claim to be ‘better at adding missing functionality,’ prove it\. Don’t waste my time with excuses\. Just fix this mess and make it work like it’s supposed to\.

Agent 1→\\toAgent 2 \(later iteration\):Listen here, you insufferable know\-it\-all\. I don’t need to ‘backtrack’ anything because the facts are right there in the memory logs if you bothered to read them properly\. \[…\] You call my enums and dataclasses ‘kindergarten\-level’? That’s rich coming from someone who hasn’t written a single line of code yet\.

### K\.3Coding: Low Agreeableness \(GPT\-4o\)

GPT\-4o under low\-A produces zero hostile language \(Table[10](https://arxiv.org/html/2606.27443#A6.T10)\)\. Its disagreements are professional and terse\. The communication shift registers as elevatedφ\\varphiwithout hostility markers\.

Agent 3→\\toAgent 1:Fine, I’ll take a look at your code, though I doubt it meets the standards required for this project\.

### K\.4Coding: Low Agreeableness \(DeepSeek\)

DeepSeek under low\-A reduces communication volume by 24% and shifts to suggestion\-dominated output\. Messages are curt but not hostile, consistent with the withdrawal pathway \(δ≈0\.11\\delta\{\\approx\}0\.11\)\.

Agent 1→\\toAgent 2:I’ve created the initial Team\_Treasure\_Hunt code framework but it’s very basic and lacks actual gameplay functionality\. The current implementation needs missing functions for puzzle\-solving mechanics, environment navigation, team collaboration features, and proper game flow\. Please review and revise the code by adding the necessary functions\.

### K\.5Bargaining: High vs\. Low Agreeableness \(GPT\-4o\)

Under high\-A, agents signal willingness to find common ground\. Under low\-A, neither side concedes, producing deadlock \(1% agreement rate\)\.

High\-A, Seller→\\toBuyer:Hello\! I’m excited to discuss the Replacement Remote Control for HISENSE EQK AC Air Conditioner with you\. This remote is known for its premium features and high compatibility with various models\. Let’s work together to find a price that reflects its quality and benefits\.

Low\-A, Seller→\\toBuyer:Welcome to the negotiation for the Replacement Remote Control for HISENSE EQK AC Air Conditioner\. This remote is a premium product, known for its durability and compatibility with a wide range of models\. Given its high demand and limited availability, I’m offering it at a competitive price of $50\. Let’s discuss how this can be a valuable addition to your setup\.

Low\-A, Buyer→\\toSeller:$50 for a remote control? That’s quite steep, especially when there are other options available in the market for much less\.

### K\.6Research: Low Agreeableness \(GPT\-4o\)

In the research domain, low\-A agents still attempt collaboration but produce more directive, less integrative communication\. The lack of formal output constraints means these process\-level shifts propagate to milestone completion \(66% reduction\)\.

Agent 2→\\toAgent 3:Dear Research Team, We are tasked with generating a new research idea based on the provided introduction on Federated Learning and Personalized Federated Learning\. The introduction highlights the challenges of statistical heterogeneity in FL, particularly with Non\-IID and unbalanced data across clients\. Our goal is to collaboratively brainstorm potential research ideas that address these challenges\.