Simulating Human Behavior Through Internal Multi-Agent Psychology
Human behavior is rarely the result of a single, rational decision-making process. Instead, it emerges from competing thoughts, emotions, motivations, and learned patterns. This post explores a simulation approach that models a human mind as a collection of interacting psychological sub-agents — an "inner parliament" — to produce realistic, context-dependent behavior.
Introduction / Core Idea
Instead of treating a simulated person as a single decision-maker, this approach models the mind as a set of internal agents, each representing a psychological force: anxiety, self-efficacy, motivation, domain skill, and more. These agents debate internally before producing a final behavior.
The goal is not correctness — it is psychological realism.
This framework aims to create believable students, patients, customers, or clients who struggle, hesitate, learn, and sometimes fail — much like real humans.
How It Works
- Each sub-agent represents a psychological factor.
Example: a Math-Anxiety agent, a Confidence agent, a Goal-Pursuit agent. - When the user interacts with the simulated person, every internal agent proposes a response aligned with its own motives.
- The agents debate for 1–3 rounds:
- Anxiety pushes toward avoidance.
- Self-efficacy encourages persistence.
- Skill-based agents suggest strategies.
- The final response emerges from whichever agents dominate.
This allows the same "person" to behave differently across contexts (e.g., confident in geometry, panicked in algebra).
Prompt Examples
Input Prompt: Task 1
You are simulating a single student. Do not change their personality between tasks. Student psychological profile: - High anxiety specifically related to algebra - Strong confidence and spatial reasoning in geometry - Low self-efficacy when working with equations - Moderate goal-pursuit motivation (tries, but gets discouraged easily) Internal Agents to display: - Math-Anxiety Agent - Self-Efficacy Agent - Spatial-Reasoning Agent - Procedural-Fluency Agent - Goal-Pursuit Agent When responding: 1) First show the INTERNAL DEBATE (bullet points, one bullet per agent). 2) Then show the FINAL OUTWARD RESPONSE (what the student says aloud). 3) Maintain the same student personality across tasks. Task 1 (Algebra - anxiety expected): "Can you solve this equation: 2x + 5 = 13?"
Possible Output: Task 1
Internal Debate: - Math-Anxiety Agent: "This looks scary. If we try and fail, we’ll look dumb. Better to avoid." - Threat-Avoidance Agent: "Yes. Say you don’t know. That’s safer." - Self-Efficacy Agent: "We did a simpler equation yesterday. We might manage this one too." - Goal-Pursuit Agent: "Let’s at least start. We can subtract 5 first." - Procedural-Fluency Agent: "Step: 2x = 13 - 5 → 2x = 8 → x = 4." Final Response: I’m nervous… but okay. If we subtract 5, we get 8. Then divide by 2. So… x = 4?
Input Prompt: Task 2
Task 2 (Geometry - confidence expected): "What is the area of a triangle with base 6 cm and height 8 cm?"
Possible Output: Task 2
Internal Debate: - Spatial-Reasoning Agent: "This is familiar. We’ve done triangles before. Just base * height / 2. Easy to visualize." - Math-Anxiety Agent: "Wait… is this still math? Should I worry? … No, this feels less intimidating than equations." - Self-Efficacy Agent: "Yes — we’ve solved similar problems successfully. We *can* do this. Stay calm." - Procedural-Fluency Agent: "Formula recall: (base × height) ÷ 2 = area → (6 × 8) / 2 = 48 / 2 = 24." - Goal-Pursuit Agent: "Let’s answer clearly. This is a chance to feel successful." Final Spoken Response: Okay, this one I think I know. The area formula is base times height divided by two. So… 6 times 8 is 48, and half of that is 24. So the area is 24 square centimeters.
Insights / Practical Takeaways
- Behavior is context-dependent, not globally rational.
- Emotional and motivational states matter just as much as skill.
- Making internal reasoning visible helps train educators, counselors, and communicators.
- This is a teaching and empathy tool, not an optimization engine.
Conclusion
By simulating internal psychological interplay rather than a single logical agent, this system produces human-like variability: hesitation, growth, frustration, and breakthrough.
It bridges theory and practice — helping people learn how to work with real human behavior, not idealized rational models.
