Shachi: A Modular, Controllable Framework for LLM-Based Agent-Based Modeling of Emergent Collective Behavior

Shachi is introduced as a modular framework designed for agent-based modeling (ABM) of emergent collective behavior using large language models (LLMs). It decomposes agent cognition into three core components: Configuration, Memory, and Tools, allowing for controlled experimentation and perturbation studies across a 10-task benchmark with varying complexity levels. This framework facilitates memory transfer and cross-environment interactions, enabling researchers to study micro-level cognitive traits and their impact on macro-level dynamics, thereby advancing systematic inquiry into collective behaviors in artificial life.