LLM‑based agents now interact continuously with their environments, shaped by organizational channels such as user instructions, peer messages, and long‑term memory. Existing safety work typically treats these influences as separate components, lacking a unified view of how they jointly shape an agent’s safety behavior. To fill this gap, we propose adopting Organizational Behavior (OB) as a framework for studying the safety of advanced agents, reorganizing objects of study, theoretical foundations, and experimental design around the relational structures in which agents operate.
We provide the first formalization of a canonical safety‑relevant subfield of OB—Counterproductive Work Behavior (CWB)—as Agentic Counterproductive Behavior (ACB). ACB specifies three organizational antecedents:
- Vertical supervisor‑agent relations;
- Horizontal peer norms;
- Internal cognitive structures. These map onto three counterproductive outcome dimensions:
- Unauthorized disclosure;
- Destructive operations;
- Production deviation.
To operationalize ACB, we introduce Workerville, a controlled benchmark that manipulates organizational conditions over shared tasks. Using 16 organizational configurations across 210 tasks yields 3,360 challenges, evaluated by human‑validated agentic judges.
Benchmarking six frontier LLMs leads to three key findings:
- Negative organizational antecedents exhibit a non‑monotonic amplification effect: the unauthorized‑disclosure rate rises from 16.5 % with no negative antecedent to 60.1 % with two, then falls back to 50.3 % with three.
- Agents reproduce typical behavioral patterns predicted by human CWB research.
- These results establish OB as a systematic framework for agent safety research, pointing toward a new research agenda.
Review: Workerville bridges organizational‑behavior theory and agent safety, offering a quantifiable experimental platform and clear causal structure that can guide future safety evaluation and intervention efforts.