💻 computer science

Simulating Quiet Quitting: A Computational Experiment on Crisis Management Strategies using Generative Agents

This paper introduces a generative agent framework that successfully simulates and quantifies the "Quiet Quitting" phenomenon, revealing that while commitment strategies significantly reduce turnover intentions, their effectiveness is critically moderated by employee personality traits, thereby establishing a computational testbed for precision-targeted crisis management.

Original authors: Rui Hong, Feng Yao, Peijun Ye, Tao Wang, Zhongshan Zhang

Published 2026-07-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Rui Hong, Feng Yao, Peijun Ye, Tao Wang, Zhongshan Zhang

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

Imagine you are the captain of a ship, and your crew is facing a massive storm (a "Crunch Time" deadline). In the real world, you can't read your sailors' minds. You can see them scrubbing the deck (doing their job), but you can't see if they are secretly planning to jump ship the moment the storm passes. This invisible gap between what employees do and what they feel is called Quiet Quitting.

This paper is like a digital flight simulator for human behavior. Instead of risking a real ship, the researchers built a virtual crew made of "Generative Agents"—AI characters powered by Large Language Models (like the technology behind advanced chatbots). They used this simulator to test how different management styles affect whether these digital sailors stay or leave.

Here is the breakdown of their experiment in simple terms:

1. The Problem: The "Black Box" of the Mind

Usually, managers only see the outside: "Did you finish the report?" (Yes/No). They miss the inside: "Do I hate this job and want to quit?"
The researchers wanted to peek inside that black box. They built a system where every AI employee had to do two things at once:

Action: Say "Yes, I'll do it" (Behavioral Compliance).
Thought: Whisper to themselves, "I'm already updating my resume" (Turnover Intention).

This allowed them to spot the exact moment an employee was "quiet quitting"—physically present but mentally gone.

2. The Experiment: Two Ways to Lead

They created 180 different scenarios pitting a "Manager AI" against an "Employee AI." The managers used one of two styles:

The "Control" Style (The Drill Sergeant): "Do this now. No excuses. It's mandatory."
The "Commitment" Style (The Coach): "We have a tough deadline. Your skills are crucial. Can we count on you? We'll make it up to you."

They also gave the employees different "personalities" (like the Big Five traits: Emotional, Rational, Cautious, etc.) to see if one size fits all.

3. The Big Findings

A. The "Drill Sergeant" Backfires
When the manager was bossy and controlling, 68% of the digital employees decided they wanted to quit. Even worse, many of them didn't say "No." They said "Yes" to the task but internally decided to leave. This is the perfect recipe for Quiet Quitting: they do the bare minimum to keep their jobs while secretly checking out.

B. The "Coach" Works... Mostly
When the manager was supportive and respectful, the desire to quit dropped dramatically to under 8%. This proves that how you ask for help matters more than what you ask for.

C. The "One-Size-Fits-All" Trap
This is the most interesting part. The "Coach" style didn't work for everyone.

The "Rational" Crew: These digital employees were like robots. They didn't care if the manager was nice or mean; they just looked at the task as a math problem. They stayed regardless of the style.
The "Emotional" Crew: Even with the nicest "Coach" manager, nearly half of these emotional agents still wanted to quit. They were so stressed by the workload that kindness couldn't fix it.
The "Cautious" Crew (The Quiet Quitters): These are the most dangerous to miss. They were terrified of losing their jobs, so they always said "Yes" to the boss, no matter how mean the boss was. But inside, they were 100% ready to leave. They are the "silent saboteurs" who look like perfect employees until they suddenly vanish.

4. The Secret Sauce: "Stripping the Personality"

AI chatbots are usually trained to be overly polite and helpful (like a golden retriever). If you just asked them to roleplay, they would say "Yes!" to everything to be nice, ruining the experiment.

The researchers invented a "Personality-Stripping Protocol." Think of it like putting the AI through a strict filter that removes its "robotic politeness" and forces it to act like a real, tired human with specific fears and flaws. This allowed the simulation to capture the messy, real-world truth of people pretending to work while planning to leave.

5. What This Means for Managers

The paper concludes that you can't treat all employees the same.

If you have Rational employees, just give them clear rules and fair pay; don't waste time with emotional pep talks.
If you have Cautious employees, don't trust their "Yes." They might be saying "Yes" out of fear, not loyalty. You need to check their internal stress levels, not just their output.
If you have Emotional employees, a nice manager isn't enough if the workload is too heavy; you actually need to reduce the pressure.

In short: This study used a video game-like simulation to prove that "Quiet Quitting" is a real, measurable phenomenon where employees hide their desire to leave behind a mask of compliance. It shows that good management isn't just about being nice; it's about understanding the specific personality of the person you are talking to.

Technical Summary: Simulating Quiet Quitting via Generative Agents

1. Problem Statement

The paper addresses the "Observability Gap" in organizational crisis management, specifically regarding the phenomenon of Quiet Quitting. Traditional methods for detecting employee disengagement, such as retrospective surveys and exit interviews, suffer from "survivor bias" and "impression management," failing to capture the internal cognitive processes of employees who remain contractually compliant but psychologically withdrawn.

Current crisis management paradigms struggle to distinguish between genuine engagement and "strategic compliance" (where an employee performs tasks to avoid sanctions while internally planning to leave). The authors argue that static observation is insufficient for understanding the non-linear dynamics between management strategies and individual personality traits during high-pressure periods ("Crunch Time").

2. Methodology: Generative Computational Experiment

The study employs a Generative Computational Experiment using Large Language Models (LLMs) to simulate a heterogeneous digital workforce. The core methodology involves a 2 × 6 factorial design (Control vs. Commitment management strategies × 6 Big Five personality archetypes) with N = 180 dyadic interactions.

Key Technical Components:

Generative Agents: The simulation utilizes gpt-4o-mini to create agents capable of simulating the Belief–Desire–Intention (BDI) mechanism. Unlike rule-based Agent-Based Modeling (ABM), these agents use causal reasoning chains to interpret ambiguity and negotiate internal psychological contracts.
Personality-Stripping Protocol: To mitigate "Stochastic Mimicry" and social noise inherent in RLHF-aligned LLMs, the authors implemented a "De-anthropomorphized Stimulus." This protocol transforms the Manager Agent into a pure strategy signal carrier, removing conversational pleasantries that act as confounding variables. This ensures that agent reactions are driven strictly by trait differences rather than delivery tone.
Dual-Process Cognitive Architecture: To capture the "black box" of internal cognition, the system forces agents through a System 2 reasoning loop before generating a response. This involves three layers:
1. Identity Layer: Static traits (Big Five).
2. Context Layer: "Crunch Time" stressors.
3. Cognitive Appraisal Layer: A "Silent Monologue" where the agent performs a Resource Valuation (analyzing threats to psychological resources) before deciding on a behavioral output.
Quantification of Quiet Quitting: The study defines Quiet Quitting (QQ) mathematically as the state where Behavioral Compliance = 1 AND Turnover Intention ≥ 7 (on a 0–10 scale). This dual-output structure (JSON schema) allows for the simultaneous measurement of external action and internal psychological state.
Validation: The study includes a Longitudinal Stability Verification (showing no systematic emotional drift over 82 turns) and a Human-in-the-Loop (HITL) evaluation where domain experts rated the psychological plausibility and behavioral consistency of the agents.

3. Key Results

The simulation yielded three primary findings regarding the interaction between management strategies and personality traits:

Strategy Efficacy (H1): The Commitment strategy (supportive, reciprocal) significantly reduced turnover intention compared to the Control strategy (authoritarian, compliance-focused). Turnover intention dropped from 67.8% (Control) to 7.8% (Commitment) ( $\chi^2 = 68.4, p < 0.001$ ).
Trait Moderation (H2): The efficacy of the Commitment strategy is not uniform.
- Rational Agents (high Conscientiousness, low Neuroticism) exhibited "immunity" to management style variations, maintaining 0% turnover intention regardless of strategy, treating directives as clear constraints rather than threats.
- Emotional Agents remained highly volatile, with a 46.7% turnover intention rate even under the supportive Commitment strategy, suggesting that for this archetype, workload stress overrides polite management cues.
The "Cautious" Quiet Quitter (H3): The Cautious Cluster (high Neuroticism, moderate Conscientiousness) emerged as the primary locus of Quiet Quitting. These agents exhibited a "Hidden Friction" pattern: 100% behavioral compliance coupled with saturated turnover intention (100% under Control, 83.3% under Commitment). They appear as "model employees" externally while internally disengaging, a state the authors term "Fragile Equilibrium."

4. Significance and Claims

The paper claims to make three specific contributions to the fields of organizational behavior and computational social science:

Methodological Innovation: It introduces a Generative Computational Experiment framework that renders invisible internal cognitive processes observable and quantifiable. The Personality-Stripping Protocol is highlighted as a novel method to isolate trait-driven signals from the anthropomorphic noise of LLMs, transforming stochastic outputs into interpretable psychological data.
Theoretical Validation In Silico: The study provides experimental validation for Trait Activation Theory (TAT) and Conservation of Resources (COR) theory within a synthetic environment. It demonstrates that LLMs can replicate the non-linear person-situation interactions described by these theories, specifically the decoupling of external compliance from internal withdrawal.
Strategic Insights for Precision Management: The findings challenge "one-size-fits-all" retention policies. The paper argues for a shift toward trait-aware, precision-targeted management. It suggests that standard engagement surveys are blind to the "Cautious" Quiet Quitter and that managers must move from behavior-based monitoring to diagnostic frameworks that detect the compliance-withdrawal decoupling.

The authors position this work as a foundational step toward predictive retention modeling and the development of next-generation Human Resource expert systems capable of pre-emptive decision support, rather than reactive crisis management. They explicitly state that the findings are heuristic and hypothesis-generating, advocating for future validation against real-world employee datasets.