Environment-free Synthetic Data Generation for API-Calling Agents

Researchers at Apple introduce ESAT, an Environment-Free Synthetic Agentic Trajectory pipeline that generates API-calling training data from API specifications alone, without executable services, databases, or human-written tasks. A GLM-4.7-FP8 task generator creates coverage-balanced, intent-level tasks; a GLM-5.1-FP8 teacher agent solves them step by step while a stateful simulator produces schema-compliant responses conditioned on the task and per-app interaction history; then Gemini-3.1-Pro from Google DeepMind filters trajectories for correctness and completeness. The authors evaluate supervised fine-tuning on AppWorld and OfficeBench, including write operations and long-horizon multi-app workflows. ESAT produces 9K AppWorld trajectories from seven real AppWorld apps, 6K trajectories from 52 synthetic apps spanning 1,017 APIs, and 1.7K OfficeBench trajectories. Fine-tuning yields gains up to 50.5% on AppWorld and 60.5% on OfficeBench, with synthetic-app data often outperforming trajectories collected in the real AppWorld environment and enabling smaller Qwen3 and Qwen3.5 models to rival larger systems such as GPT-4o and Nemotron-3-120B. Quality analysis using OpenAI’s GPT-5.1 finds 93.7% of simulated API responses valid, while Gemini-3.1-Pro achieves 95.2% precision against AppWorld’s executable verifiers. The central result is that an LLM can serve as a persistent digital world model, making scalable, stateful API-agent supervision a specification problem rather than an infrastructure project.