AutoEnvScaling Lets Agents Build Their Own Training Environments

Researchers led by Natasha Jaques introduce AutoEnvScaling, where a proposer agent autonomously builds Harbor task environments and sandboxes for RL self-play, minimizing human involvement in recursive self-improvement. The approach boosts Qwen3.6-35B's Terminal Bench score from 41.1 to 53.3.

2026-10-14 ~ 2026-10-15 · 3 related posts

• AutoEnvScaling: agents build their own training tasks for human-free self-play RL — _AndrewZhao · 2026-10-14
• AutoEnvScaling: Agents Build Their Own Post-Training Environments, Terminal-Bench 41.1 to 53.3 — rbhar90 · 2026-10-14
• AutoEnvScaling: proposer agent builds its own RL environments, Qwen3.6-35B doubles Terminal Bench 4 score — jzl86 · 2026-10-15