Generating synthetic data to train and evaluate LLM agents in enterprise contexts has long been a compromise: either use real database schemas with legal and security risks, or adopt statistical approaches that fail without seed data. SAP has chosen a third path: simulate the enterprise environment itself.

The Synthesis Through Simulation (STS) framework, now open source on GitHub, flips the paradigm. Instead of extracting data from real systems, an LLM agent executes operations against policy-enforcing APIs within simulated environments. Because data is generated through the same environment that defines what is valid, STS guarantees structural validity by construction. The key lies in decoupling constraint enforcement from distribution modeling: two problems that previously traveled together, often at the expense of one or the other.

The central component, Generalist Populator (GP), is a domain-agnostic agent that reached an average marginal fidelity of 0.88 and 100% constraint satisfaction across ten environments, without access to database schemas. Statistical synthesizers are inapplicable in seven of these environments because they require seed data, while schema-privileged agents fail 82% of trajectories in the airline environment due to brittle task composition.

For those evaluating on-premise deployment, the message is structural. The ability to generate synthetic data coherent with enterprise constraints without exposing or copying the real schema changes the data sovereignty calculus. There is no longer a need to move sensitive datasets to the cloud to train specialized agents: one can build a local pipeline that starts from simulation and arrives at fine-tuning, maintaining full control. This reduces dependence on external vendors and shifts attention to the quality of the simulated environment, which becomes the new critical point.

The STS approach also signals a shift in incentives for enterprise agent developers. Teams managing self-hosted infrastructure can iterate on synthetic datasets without waiting for legal approval to use real data, accelerating development cycles. At the same time, cloud data generation providers must contend with the possibility that companies prefer local solutions when compliance demands it.

Limits remain: the marginal fidelity of 0.88 indicates that some distributions remain approximate, and simulation quality depends on the fidelity of the policies implemented. But the open-sourcing of the code and ten environments allows the community to verify and adapt the framework to their own domains, a concrete step toward more reliable LLM agents in regulated contexts.