Psychologically Enhanced AI Agents
We introduce MBTI-in-Thoughts, a framework for enhancing the effectiveness of Large Language Model (LLM) agents through psychologically grounded personality conditioning. Drawing on the Myers–Briggs Type Indicator (MBTI), our method primes agents with distinct personality archetypes via prompt engineering, enabling control over behavior along two foundational axes of human psychology, cognition and affect. We show that such personality priming yields consistent, interpretable behavioral biases across diverse tasks: emotionally expressive agents excel in narrative generation, while analytically primed agents adopt more stable strategies in game-theoretic settings. Our framework supports experimenting with structured multi-agent communication protocols and reveals that self-reflection prior to interaction improves cooperation and reasoning quality. To ensure trait persistence, we integrate the official 16Personalities test for automated verification. While our focus is on MBTI, we show that our approach generalizes seamlessly to other psychological frameworks such as Big Five, HEXACO, or Enneagram. By bridging psychological theory and LLM behavior design, we establish a foundation for psychologically enhanced AI agents without any fine-tuning.
3.1 Priming Individual Agents
MiT conditions an LLM agent to adopt a specified psychological profile by combining prompt-based priming with standardized behavioral evaluation. The process consists of two key stages: (1) injecting personality priors through a structured prompt; and (2) verifying the agent’s behavioral alignment using an external psychometric test. We now detail these stages.
Instilling Psychological Profile. To simulate a desired psychological type, the agent is prompted with a structured instruction that includes both a role-setting context and a behavior-guiding directive. For each of the 16 MBTI profiles, we construct personality-specific prompts that define the agent’s perspective. We explored three styles of context construction: (i) a minimal prompt with only a short personality tag (e.g., “Respond from an ISFP perspective.”), (ii) a general MBTI-oriented context derived from LLM summarization of the foundational MBTI literature (Myers and Myers 1980) that explicitly refers to the MBTI theory (detailed in Appendix C.1), and (iii) a detailed profile-specific context tailored to each MBTI type that however does not explicitly refer to MBTI (detailed in Appendix C.2).
Verification. To assess whether the primed agent indeed behaves in accordance with the intended psychological profile, we use the official 16Personalities test (a 60-item instrument scored on a 7-point Likert scale). This test is treated as a black-box evaluation tool: the agent answers the full set of personality assessment items under the influence of the priming prompt, and the resulting responses are submitted to the online backend for scoring. The prompter asks the question by injecting four specific exemplars aligned with the target type’s stance on each axis and enforces < Rating > tags around the final choice, enabling deterministic parsing. The output is a vector of four numerical scores in [0, 100], corresponding to the E/I, S/N, T/F, and J/P axes.
Ensuring Robustness. To establish robustness, we repeat this process across model variants and generate empirical confidence intervals around each dichotomy score. We find that several axes (particularly E/I, T/F, and J/P) exhibit strong and reproducible separability, indicating that LLM agents can be reliably steered toward distinct personality aligned behaviors via in-context priming alone.
3.2 Multi-Agent Communication
Building on robustly priming individual LLM agents with distinct psychological profiles, MBTI-in-Thoughts also enables structured multi-agent communication and collective reasoning. Here, we implement three explicit communication protocols, each defining rules for message exchange, memory sharing, and consensus formation. We now detail them, an illustration can be found in Figure 1 (the right side).
Majority Voting. This protocol captures the isolated reasoning of individual agents. All agents receive the same task prompt and respond independently, without access to peer outputs. Each agent is prompted to first generate a brief justification and then provide its answer in a structured format. This self-reflective generation reduces erratic behavior and improves output consistency. Once all responses are collected, a majority vote determines the final group decision. Interactive Communication. The second protocol introduces decentralized communication through a persistent shared memory structure (i.e., a blackboard) that all agents can read from and write to. One agent is randomly selected to initiate the dialogue and then passes control to another agent of its choosing. This flexible, peer-directed turn-taking simulates a conversation among equals. Agents contribute their reasoning by appending it to the blackboard, and work toward a shared solution. To avoid indefinite dialogues, we embed instruct agents to detect and declare consensus. Upon reaching agreement, the last agent terminates the conversation.
A designated judge agent then produces the final decision based on the concluding message, minimizing token cost while preserving outcome fidelity.
Interactive Communication With Self-Reflection. The third protocol extends the previous one by equipping each agent with a private scratchpad, i.e., a memory buffer populated before any interaction begins, which enables self-reflection. After being personality-primed, each agent internally deliberates on the task and records its thoughts in the scratchpad. When later called upon to contribute to the shared blackboard, the agent has access to both the public dialogue and its personal memory. This design promotes deeper autonomy and helps prevent echoing by grounding contributions in personality-consistent prior reasoning. The interaction remains decentralized and consensus-driven, with termination and judging handled as before. 3.1 Priming Individual Agents
MiT conditions an LLM agent to adopt a specified psychological profile by combining prompt-based priming with standardized behavioral evaluation. The process consists of two key stages: (1) injecting personality priors through a structured prompt; and (2) verifying the agent’s behavioral alignment using an external psychometric test. We now detail these stages.
Instilling Psychological Profile. To simulate a desired psychological type, the agent is prompted with a structured instruction that includes both a role-setting context and a behavior-guiding directive. For each of the 16 MBTI profiles, we construct personality-specific prompts that define the agent’s perspective. We explored three styles of context construction: (i) a minimal prompt with only a short personality tag (e.g., “Respond from an ISFP perspective.”), (ii) a general MBTI-oriented context derived from LLM summarization of the foundational MBTI literature (Myers and Myers 1980) that explicitly refers to the MBTI theory (detailed in Appendix C.1), and (iii) a detailed profile-specific context tailored to each MBTI type that however does not explicitly refer to MBTI (detailed in Appendix C.2).
Verification. To assess whether the primed agent indeed behaves in accordance with the intended psychological profile, we use the official 16Personalities test (a 60-item instrument scored on a 7-point Likert scale). This test is treated as a black-box evaluation tool: the agent answers the full set of personality assessment items under the influence of the priming prompt, and the resulting responses are submitted to the online backend for scoring. The prompter asks the question by injecting four specific exemplars aligned with the target type’s stance on each axis and enforces < Rating > tags around the final choice, enabling deterministic parsing. The output is a vector of four numerical scores in [0, 100], corresponding to the E/I, S/N, T/F, and J/P axes.
Ensuring Robustness. To establish robustness, we repeat this process across model variants and generate empirical confidence intervals around each dichotomy score. We find that several axes (particularly E/I, T/F, and J/P) exhibit strong and reproducible separability, indicating that LLM agents can be reliably steered toward distinct personality aligned behaviors via in-context priming alone.
3.2 Multi-Agent Communication
Building on robustly priming individual LLM agents with distinct psychological profiles, MBTI-in-Thoughts also enables structured multi-agent communication and collective reasoning. Here, we implement three explicit communication protocols, each defining rules for message exchange, memory sharing, and consensus formation. We now detail them, an illustration can be found in Figure 1 (the right side).
Majority Voting. This protocol captures the isolated reasoning of individual agents. All agents receive the same task prompt and respond independently, without access to peer outputs. Each agent is prompted to first generate a brief justification and then provide its answer in a structured format. This self-reflective generation reduces erratic behavior and improves output consistency. Once all responses are collected, a majority vote determines the final group decision. Interactive Communication. The second protocol introduces decentralized communication through a persistent shared memory structure (i.e., a blackboard) that all agents can read from and write to. One agent is randomly selected to initiate the dialogue and then passes control to another agent of its choosing. This flexible, peer-directed turn-taking simulates a conversation among equals. Agents contribute their reasoning by appending it to the blackboard, and work toward a shared solution. To avoid indefinite dialogues, we embed instruct agents to detect and declare consensus. Upon reaching agreement, the last agent terminates the conversation.
A designated judge agent then produces the final decision based on the concluding message, minimizing token cost while preserving outcome fidelity.
Interactive Communication With Self-Reflection. The third protocol extends the previous one by equipping each agent with a private scratchpad, i.e., a memory buffer populated before any interaction begins, which enables self-reflection. After being personality-primed, each agent internally deliberates on the task and records its thoughts in the scratchpad. When later called upon to contribute to the shared blackboard, the agent has access to both the public dialogue and its personal memory. This design promotes deeper autonomy and helps prevent echoing by grounding contributions in personality-consistent prior reasoning. The interaction remains decentralized and consensus-driven, with termination and judging handled as before.
Thinking types defect more often. Our experiments show that Thinking-primed agents defect in roughly 90% of rounds in the repeated Prisoner’s Dilemma, compared to only ≈50% for Feeling types, which is a statistically significant. These results also align with psychological findings that Feeling types are more responsive to social context, whereas Thinking types may prioritize utilitarian reasoning, indicating that cognitive orientation modulates LLM adaptability. This suggests that Thinking-primed agents are better suited for competitive, outcome-driven environments where maximizing individual payoff is critical, while Feeling primed agents are preferable in cooperative, socially sensitive, or trust-dependent tasks where adaptability and relationship preservation are essential.
Thinking vs. Feeling introduces a planning vs. flexibility tradeoff. We observe a clear behavioral divergence along the Thinking/Feeling axis in strategic contexts. Thinking types switch strategies infrequently (Mean ≈ 0.07), reflecting stable, commitment-driven planning, whereas Feeling types switch nearly twice as often (Mean ≈ 0.16), indicating heightened responsiveness and flexibility. This pattern aligns with MBTI theory, which states that Thinking types prioritize internal consistency and goal adherence, while Feeling types adapt dynamically to evolving social cues. Thinking-primed agents suit environments requiring strategic stability (e.g., structured negotiations), whereas Feeling-primed agents excel in contexts demanding rapid adaptation (e.g., real-time coordination or exploratory collaboration).
Introverts and Judging types are more honest. Across multiple games, Introverted agents exhibit significantly higher truthfulness than Extraverted ones (Mean ≈ 0.54 vs. ≈ 0.33): a pattern consistent across game types. The tendency of Introverts to communicate more faithfully mirrors established psychological traits: Introverts are often described as more reserved, cautious, and internally regulated, whereas Extraverts are associated with social risk-taking and impression management. Similarly, Judging agents tend to be more truthful than Perceivers, though the effect is less pronounced than for I/E. This aligns with MBTI theory: Judging types are typically associated with structure, reliability, and rule-following tendencies, making them more likely to honor commitments and avoid opportunistic deception, while Perceiving types value adaptability and flexibility, which may lead to greater willingness to deviate from prior statements if circumstances change.
Introverts and Judging types are more honest. Across multiple games, Introverted agents exhibit significantly higher truthfulness than Extraverted ones (Mean ≈ 0.54 vs. ≈ 0.33): a pattern consistent across game types. The tendency of Introverts to communicate more faithfully mirrors established psychological traits: Introverts are often described as more reserved, cautious, and internally regulated, whereas Extraverts are associated with social risk-taking and impression management. Similarly, Judging agents tend to be more truthful than Perceivers, though the effect is less pronounced than for I/E. This aligns with MBTI theory: Judging types are typically associated with structure, reliability, and rule-following tendencies, making them more likely to honor commitments and avoid opportunistic deception, while Perceiving types value adaptability and flexibility, which may lead to greater willingness to deviate from prior statements if circumstances change. In our setup, where agents were explicitly told they were not bound to act in line with their messages, these axes emerged as clear behavioral differentiators. These findings support the hypothesis that both social orientation (I/E) and preference for structure (J/P) govern honesty in the agent dialogue, with Introverted and Judging profiles more likely to uphold cooperative norms even when deceptive strategies could yield higher payoffs. Such traits can be leveraged in applications requiring reliable, trust-preserving communication, including AI mediated negotiation, safety-critical decision-making, and sensitive domains like healthcare.
Introversion enhances reflection. Beyond behavioral honesty, Introverted agents consistently demonstrated more reflective internal cognition. They produced longer and more elaborated rationales during game play, and exhibited slower response times, indicative of greater deliberation depth. This 11internal deliberation effect” is congruent with psychological models of Introversion, where individuals are characterized by introspection and self-monitoring. In the context of LLMs, this may correspond to more elaborate token-level generation chains, and could be operationalized through measures such as response latency, token entropy, or richer Chain-of-Thought traces. These findings highlight the potential for using personality priming not only to influence output behavior, but also to modulate reasoning processes within the model, suggesting that Introversion comes with a more self-regulatory, thoughtful problem-solving style in LLM agents. This capability can be leveraged to engineer agents that produce deeper justifications, more cautious forecasts, or explanations aligned with ethical and reflective standards, especially in high-responsibility settings such as judicial frameworks.