Apple trains API-calling agents with synthetic trajectories and no live environment
apple · hf · 2026-07-21
### What Apple proposes Apple introduces an **environment-free synthetic data** pipeline for training API-calling agents. ### How it works - Given only API specs, an LLM generates tasks that are solvable with those APIs. - A teacher agent solves the task while an LLM simulator produces coherent synthetic API responses conditioned on task context and interaction history. - An LLM judge filters low-quality trajectories. ### Why it matters - Collecting agent trajectories usually requires fully implemented environments, executable APIs, and realistic backend data. - This method removes that bottleneck and lets teams generate supervision without a live environment. - Fine-tuning on the synthetic data improves performance on **AppWorld** and **OfficeBench**, including both retrieval-style and state-changing tasks.
More from coding & agent
- The author says Codex reached 20x and is now debugging spec decoding on a hybrid parallel setup — TheZachMueller · 2026-07-21
- Axcess adds an MCP connector for WCAG accessibility checks that scanners miss — modelcontextprotocol · 2026-07-21
- X post asks whether Cursor Composer, built on Kimi models, would also be banned — max_paperclips · 2026-07-21
- A developer’s Codex usage is draining pooled enterprise credits at a small company — Distinct_Relation_62 · 2026-07-21
- Qwen Code ships cua-driver-rs 0.7.3 with relative coordinates and MCP filtering — github-actions[bot] · 2026-07-21
- Matt Pocock says every new codebase turns legacy within days — mattpocockuk · 2026-07-21