Evolving terminal worlds lift agent scores 18 points

A new Environment Evolution paper makes terminal-agent training worlds harder off-policy and reports 14.4- and 18.0-point gains on Terminal-Bench 2.1 for Qwen3.6-27B and Qwen3.6-35B-A3B after long-horizon reinforcement learning. The method’s harness also tested against Hy4 preview, Claude Opus 5, and GPT-5.6 Sol, but the result is still a two-model paper result, not a deployment guarantee. Builders should file it beside Octobench’s evidence that harness design changes same-model outcomes and measure environment generation, reward stability, and transfer before adding this training loop.