Apple researchers have proposed a way to train API-calling agents without first building full executable test environments. The method uses large language models as on-the-fly simulators that generate tasks, synthetic API responses, and filtered training trajectories from API specifications.
The problem is that realistic agent training usually needs working APIs, backend databases, and repeatable environments. That setup is expensive to create and hard to scale across many software ecosystems. Apple’s approach asks a model to generate solvable tasks, lets a teacher agent attempt them, and uses another model to simulate coherent stateful responses.
An LLM judge then filters the resulting trajectories before they are used for fine-tuning. Apple says evaluations on AppWorld and OfficeBench showed performance gains for both information-retrieval and state-changing tasks.
The work is research, not a product release, and synthetic environments can miss real-world edge cases. Still, it points to a practical route for training software agents when companies have API documentation but cannot cheaply produce thousands of safe, realistic interactions with live systems.