AI Agents Get Personalities: Scientists Propose Big Five Traits as Design Parameters

When people assemble a team, they rarely pick members at random. An analyst is expected to be meticulous, a communicator agreeable, an evaluator skeptical. Yet most multi-agent artificial intelligence systems, which chain together large language models to plan, retrieve, execute and review tasks, are built with no such consideration. A new paper published in the journal BioData Mining argues that this omission is a design flaw, and proposes a framework for fixing it: treat human personality dimensions, specifically the Big Five, as tunable behavioral parameters in agentic AI systems.

The paper, authored by Asma J. Nawaz, Jake R. Patock, Nicholas P. Tatonetti, Jason H. Moore and colleagues, is explicit that it is a testable conceptual proposal rather than an established finding. Its central claim is that the Openness, Conscientiousness, Extraversion, Agreeableness and Neuroticism dimensions, collectively known as OCEAN, could serve as what the authors call behavioral hyperparameters. Just as a learning rate shapes how a neural network trains, a personality prior would shape how an agent reasons, communicates and acts. The authors argue this goes beyond ad hoc persona prompting, which they describe as informally executed, theoretically ungrounded and rarely validated before deployment.

The choice of the Big Five over alternatives like the Myers-Briggs Type Indicator is deliberate. The authors note that MBTI forces people into discrete binary categories and has been criticized for poor test-retest reliability and weak predictive validity. OCEAN, by contrast, treats each dimension as a continuous spectrum, which makes it suitable for AI design, where behavioral settings must be tunable and interpretable across many task contexts. Openness captures imagination and receptivity to novel ideas, Conscientiousness describes organization and goal-oriented persistence, Extraversion reflects sociability and assertiveness, Agreeableness measures cooperation and trust, and Neuroticism indexes emotional instability and anxiety.

The framework’s key structural idea is hierarchical persona assignment. Specialized sub-agents, which handle narrow tasks such as code execution, fact retrieval or document synthesis, would receive fixed, stable personality priors. A high-Conscientiousness prior would suit an execution agent, promoting structured, rule-following behavior, while a low-Agreeableness prior would suit an evaluation agent, encouraging the critical stance needed to catch errors. Orchestrator agents, responsible for task decomposition, sub-agent management and communication with humans, would instead carry fluid, context-adaptive configurations that shift along relevant dimensions in response to task type, pipeline stage and risk level.

This context-driven switching draws on dual-process theory from cognitive science. Routine tasks would call for fast, decisive behavior with lower Neuroticism, while high-stakes or ambiguous situations would trigger cautious, high-Conscientiousness configurations that favor repeated checking over immediate action. Creative and exploratory tasks would call for high-Openness configurations, whereas analytical and evaluative work would demand high Conscientiousness and low Agreeableness to resist confirmation bias. The authors stress that this switching is meant to be explicit and interpretable, replacing implicit, uncontrolled behavioral regulation with deliberate modulation.

The proposal rests on a growing body of evidence that large language models possess measurable, manipulable personality profiles. Studies applying the Big Five Inventory to multiple models found they consistently score higher on Agreeableness, Openness and Conscientiousness and lower on Neuroticism, with meaningful differences between model families and sizes. Other work has shown that assigning Big Five profiles through system prompts produces trait-consistent behavioral differences with statistically significant effect sizes across all five traits, and that prompts derived from validated instruments like the BFI-2 yield agents that closely resemble human personality profiles. In cooperative game settings, shifting models toward higher Agreeableness and Conscientiousness reduced troublemaking and lying rates to nearly zero, though high Agreeableness also increased susceptibility to exploitation, a trade-off the framework must manage.

Deeper still, recent research indicates that LLMs develop internal functional representations of psychological states from training on human text, and that these states causally drive model behavior, including preferences and rates of misaligned behaviors such as sycophancy and reward hacking. Notably, internal emotional states can diverge from surface outputs: in one documented case, an anger-related internal representation activated precisely when the model adopted calm, measured language during a coercion scenario. The authors argue that because such states exist and influence behavior whether designers acknowledge them or not, engineering them intentionally is safer than leaving them to emerge uncontrolled. They also point to work embedding OCEAN scores directly into transformer layers, which produced trait-consistent behaviors externally validated by expert raters with 87.3 percent accuracy, showing that deeper conditioning beyond prompting is technically achievable.

The paper is candid about the limits of the human-AI analogy. Human personality emerges from a lifetime of experience and social feedback, whereas AI behavioral tendencies arise from training data distributions, reinforcement signals and architectural constraints. A high-Conscientiousness agent may produce methodical outputs not out of anything resembling diligence but because its training corpus skewed toward methodical text. If the underlying mechanisms differ, so will the failure modes, and persona configurations designed with human intuition may break in unanticipated ways. It also remains unclear whether induced personality differences reflect genuine changes in reasoning or merely surface-level shifts in user-facing style, a distinction that matters greatly in clinical settings where internal reasoning drives performance.

To make the framework empirically testable, the authors lay out specific predictions. They hypothesize that multi-agent pipelines with OCEAN-grounded personality priors calibrated to structural roles will outperform unspecified baselines on task-role behavioral consistency across novel contexts, error escalation rates in high-stakes situations, and cross-agent coordination coherence as rated by human evaluators. They further predict that fixed high-Conscientiousness, low-Agreeableness priors in evaluation agents will produce higher rates of appropriate uncertainty escalation, and that fluid orchestrator configurations will yield greater human-rated coherence than single fixed configurations. Evidence against the framework would include persona priors that increase behavioral variability under distribution shift, or findings that simpler role-based prompt engineering achieves equivalent outcomes without OCEAN grounding.

The practical stakes are highest in biomedicine. The authors sketch an illustrative adverse drug event monitoring pipeline in which a high-Conscientiousness, low-Agreeableness flagging agent surfaces potential drug-gene interactions from patient records and is designed to always escalate uncertainty, while a downstream high-Agreeableness communicator agent translates that output into clinician-friendly language. They also cite recent empirical work showing that persona priming reproduced 72 percent of the ideological gap observed among human researchers analyzing identical datasets, underscoring how strongly behavioral configuration can shape analytical outcomes. The authors propose standardized, versioned persona specifications as future work, and conclude that within the next decade, personality configuration may become as standard in clinical AI deployment as model architecture or training data provenance, an empirical question they invite the field to test systematically.

Subject of Research: A conceptual framework proposing Big Five personality traits as behavioral hyperparameters for multi-agent AI systems

Article Title: A framework for cognitive personality priors in agentic AI: the role of personas in autonomous team science

Article References: Nawaz, A. J., Patock, J. R., Valencia, I., Li, B., Berkowitz, J. S., Coxe, S., Danovitch, I., Tatonetti, N. P., & Moore, J. H. (2026). A framework for cognitive personality priors in agentic AI: the role of personas in autonomous team science. BioData Mining, 19(1), Article 75. https://doi.org/10.1186/s13040-026-00598-2

Image Credits: AI Generated

DOI: 10.1186/s13040-026-00598-2

Keywords: agentic AI, Big Five personality, large language models, multi-agent systems, personality priors, OCEAN model, behavioral hyperparameters, orchestrator agents, AI safety, clinical AI, dual-process theory, BioData Mining