API training called large language modeling (LLM) requires large amounts of high quality. However, collecting such data at scale often requires fully deployed environments with usable APIs and physical, pre-populated data, creating a major scaling bottleneck. To overcome this, we propose an artificial data generation method using LLMs as models in the digital world. Given only an API specification, our method generates trajectories that simulate the interaction between an agent and an ideal environment. Specifically, LLM first generates various tasks that can be solved by the provided APIs. The teacher agent then solves each task iteratively while the LLM simulator generates the corresponding synthetic API responses conditioned on the task and simulation history. Finally, the LLM judge filters the trajectories to ensure the quality of the resulting dataset. We test our approach to the AppWorld and OfficeBench challenge, which includes both information retrieval and transformation tasks. Fine-tuning models on our synthetic data brings significant performance benefits, demonstrating that effective monitoring of API-calling agents can be done without any executable environment. Our results establish LLM-based API simulation as a viable, scalable solution for training agents across diverse ecosystems.
