Synthesis Through Simulation: Generating Coherent Enterprise Data via Scalable Agent-System Interaction

TEXT START: Tool-calling agents have become central to enterprise AI, yet training and evaluating them at scale remains severely constrained due to business and legal restrictions on enterprise systems, data, and database schemas.

THE DISSECTION

STS is an industrial workaround for the enterprise-data bottleneck. It replaces access to real enterprise systems with simulated environments whose policy-enforcing APIs make valid state transitions available by construction. The Generalist Populator then generates tool-calling trajectories without database schemas.

The reported results—0.88 average marginal fidelity, 100% constraint satisfaction across ten environments, and 82% trajectory failure for schema-privileged agents in the airline environment—show that governed approximations of enterprise operations can be manufactured at scale. The paper turns “we cannot train on the firm” into “we can simulate the firm.”

THE CORE FALLACY

The central technical error is confusing validity inside a simulation with equivalence to a real enterprise. Constraint satisfaction proves that the simulator’s rules were obeyed. It does not prove that those rules capture hidden state, rare cases, incentives, exceptions, institutional disorder, or real joint distributions. Marginal fidelity is not workflow fidelity.

The deeper Discontinuity Thesis error is treating this capability as neutral infrastructure. It is not. By reducing data, schema, and legal friction, STS makes enterprise agents cheaper and easier to train and evaluate. That strengthens P1, undermines human-only economic domains under P2, and accelerates P3. “Schema-free” is not a human moat. It is a deployment enabler.

HIDDEN ASSUMPTIONS

• The simulated policies adequately represent real enterprise rules and edge cases.
• Marginal fidelity is sufficient despite tightly coupled workflows.
• Synthetic trajectories transfer to live systems with changing policies and ambiguous goals.
• Constraint satisfaction correlates with useful operational performance.
• Simulation can substitute for proprietary reality without creating major distributional blind spots.
• The gains from scalable agents will be distributed rather than captured by owners and controllers of AI capital.

SOCIAL FUNCTION

Primary classification: transition management. Secondary classifications: partial truth and elite self-exoneration.

The paper solves a real engineering bottleneck. Its social effect is more consequential: it reframes agent-driven labor substitution as a data-generation and compliance problem. That lets institutions build the machinery of automation while avoiding the ownership question. Human displacement becomes an implementation milestone instead of a political rupture.

THE VERDICT

STS is not a defense of the post-WWII economic order. It is a logistics upgrade for its successor. If its assumptions hold, it lowers the cost and legal friction of training enterprise agents, making the severance of labor from production easier to execute.

The technical result is real but bounded: scalable simulation of constrained trajectories in ten environments, not proof of equivalence to living enterprises. Under the DT lens, STS is acceleration infrastructure—valuable to Sovereigns, useful to Servitors who control verification, orchestration, energy, logistics, or maintenance, and corrosive to routine cognitive labor. The more successful it becomes, the less mass employment remains a stable foundation.