Tool‑calling agents have become central to enterprise AI, yet scaling their training and evaluation is hampered by business and legal restrictions on systems, data, and database schemas. Tabular data synthesis offers a natural alternative but suffers from structural validity and schema availability limits; procedure‑based methods often lack distributional fidelity and require per‑domain authoring. We introduce Synthesis Through Simulation (STS), a schema‑free data synthesis paradigm where a large language model (LLM) agent generates data by invoking policy‑enforcing APIs within simulated enterprise environments. Because generation occurs in the same environment that defines validity, STS guarantees structural validity by construction and decouples validity enforcement from distribution modeling, allowing each to be tackled independently. To address distributional fidelity and scalability, STS provides the Generalist Populator (GP), a domain‑agnostic agent. GP achieves an average marginal fidelity of 0.88 and 100% constraint satisfaction across ten environments without accessing DB schemas. In contrast, statistical synthesizers are inapplicable to seven environments due to seed‑data requirements, and schema‑privileged agents fail 82% of trajectories in the airline environment because of tightly coupled workflows. The full framework, all ten environments, and generated datasets are open‑sourced at https://github.com/SAP/synthesis-through-simulation.
Review