TableFactory: Generating Semantically Linked Tabular Data via Multi-Agent Behavioral Simulation

Multi-table operational data is scarce, and much synthetic tabular work targets single-table distribution matching. We present TableFactory, a prompt-conditioned multi-agent simulator that produces semantically linked relational tables for small-business scenarios. From a short text prompt (e.g., ``simulate a butcher shop in a morning market in Barcelona''), an LLM-backed generator builds a full business configuration, such as catalogs, suppliers, staffing, demand, and policies, augmented with macro attributes such as GDP and tax rates. Five abstract entity roles interact through a shared relational world state, materializing a fixed schema of twenty foreign-key-linked tables per run. The environment supports LLM, RL (PPO, SAC, DDPG), and rule-based controllers, together with an MBTI-style persona module for behavioral diversity. We report diagnostic studies of controller interchangeability, scenario and regional variation, and persona effects, and release reproducible table bundles with configurations and metadata. We will fully open-source the codebase to support future research.