Recursive Synthesis Creates Thousands of Long-Horizon AI Tasks.

Generating high-quality, long-horizon training data for terminal agents is notoriously expensive and time-consuming, often costing hundreds to thousands of dollars per task due to the need for consistent instructions, environments, solutions, and verifiers. Human authoring struggles to scale, and direct generation by large language models (LLMs) frequently fails to maintain these critical dependencies.This research presents Recursive Synthetic Terminal Tasks (RST), a novel framework designed for the scalable and verified synthesis of long-horizon terminal-agent tasks. RST begins with verified seed tasks and then recursively extends their reference solutions, realigns the verifier and instructions, and validates the new workflow in a sandbox. Accepted tasks are then reused as seeds for subsequent rounds.Over fifteen recursive rounds, RST successfully produced 37,484 synthesized tasks at approximately $0.05 per task. The difficulty of these tasks increased substantially with each round, with reference solutions growing significantly in length and complexity. The synthesized data proved highly effective for training: supervised fine-tuning and agentic PPO on these trajectories led to significant performance improvements (up to 10 points and 41.2% relative gains) for Qwen3.5 models on challenging terminal benchmarks. The process showed no signs of ceiling, indicating its potential for even greater scale.