Apple ML proposes synthetic data method for API agents

Training API-calling LLM agents requires high-quality trajectories but typically demands fully implemented environments with executable APIs. Apple's method generates synthetic data without those environment dependencies.