Simulation without databases: How corporate agents overcome the dilemma of sensitive systems

Enterprise automation today faces a thorny paradox: software agents capable of invoking tools have become the backbone of transformation, but training and evaluating them at scale run into a solid legal and commercial wall that blocks access to operational data and real database schemas.When enterprises lock down their data vaults to protect confidentiality, models fall into the trap of fragile synthetic environments.The usual result is that automation projects stumble before they even touch the actual production environment.

A new research paper published by scholars on the Arkhive platform, led by Yiping Li and Ashutosh Hathidara, examined the failures of conventional methods for generating synthetic data. The results showed that statistical data-generation tools for tabular data failed to operate in seven out of ten environments because they lacked the primary data that feed them, while procedural approaches suffered from the absence of a natural statistical distribution without writing manual rules for each domain separately. The most striking surprise was that giving agents direct access to database schemas did not protect them from collapse; agents with schema visibility recorded a failure rate of 82 percent of paths in a complex airline-booking environment, due to the fragility of task composition and sequencing.

The solution proposed by the researchers is a model they called simulation-based synthesis, a path completely free of any need for database schemas. The model relies on a linguistic agent that generates data by executing operations directly on APIs that enforce operational policies within simulated enterprise environments. Because the data are generated inside the same environment that determines what is correct and what is incorrect, structural validity becomes guaranteed by design, with strict rule-enforcement mechanisms separated from the statistical-distribution modeling of the data.Data validity here does not come from guessing tables, but from direct interaction with business logic and system rules.

The developed agent, which is domain-independent, demonstrated its ability to solve both statistical-accuracy and scalability challenges; it achieved a 100 percent success rate in satisfying constraints across all ten environments without any access to database schemas, while also attaining an average marginal accuracy of 0.88. The team released the full framework, the ten environments, and the generated datasets as open-source projects, breaking the monopoly of closed-lab tools for evaluating agents.

This shift has a direct impact on technology teams in the Gulf and Egypt, where personal data-protection and banking-confidentiality laws in sectors such as banking, aviation and telecommunications act as a barrier that stalls independent-agent testing for months within legal-review corridors. The proposal changes the decisions of CTOs and budgeting rules, as it allows local teams to test tool-invoking agents and evaluate their behavior through simulated interfaces without having to extract sensitive customer data or wait for lengthy regulatory approvals. It also moves the required skill within regional teams from building manual database masks to designing precise simulation environments for APIs, reducing experiment costs and securing automation pipelines before they are deployed in sensitive work environments.