Environment supply limits terminal agent RL. Having agents build environments from the solver's own rollouts cuts per-environment cost and measurably lifts Terminal-Bench scores.
Key takeaways · AI-distilled
AutoEnvScaling treats environment design itself as a terminal task: a proposer AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → builds new RL environments from the solver's rollouts, and the solver's new rollouts steer the next round.
When one model, Qwen3.6-35B-A3B, played both proposer and solver, the authors report a 12.2-point gain on Terminal-Bench 2.1 and a Terminal-Bench 4.0 score rising from 1.6 to 3.2.
The 4.3-point Terminal-Bench 2.1 gain for Qwen3.5-35B-A3B used environments built by GPT-5.6-sol, measured against a matched Tmax baseline. The pipeline also generated environments for computer use, professional work and GPU kernel programming.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
articleSPROUT: Scaling Agentic Reinforcement Learning for Long-Horizon Software EngineeringMengyang Liu, Weiming Wang, Haixin Nan, Yu Wu, Peize Li, Shihong Huang, Xuning He, Xue Jiang, Yihong Dong
articleAgentic Data Engineering for LLMs: System Design and a Controlled Empirical StudyRuomeng Ding, Qianli Shen, Zhaoyang Han, Meixin Chen, Daoyuan Chen, Yaliang Li