The paper introduces environment evolution, a method for generating training environments for terminal agents that become progressively harder without requiring on-policy rollouts. Unlike previous co-evolution approaches that depend on the agent's current weaknesses, environment evolution derives three difficulty directions—scenario novelty, skill rarity, and execution length—from a multi-turn learning objective and implements them through a multi-agent harness. This allows environments to be evolved off-policy, providing continuous learning signals even as the agent improves.
Experiments with frontier models (Hy4 preview, Claude Opus 5, GPT-5.6 Sol) show that evolved environments are consistently more challenging. When used for reinforcement learning training on Qwen3.6-27B and Qwen3.6-35B-A3B, the method improves performance on Terminal-Bench 2.1 by 14.4 and 18.0 percentage points, respectively, outperforming co-evolution and ensemble baselines.
The key interest is that environment evolution decouples difficulty generation from the training agent's policy, enabling longer-lasting learning signals and better generalization. This addresses a critical bottleneck in scaling interactive environments for capable terminal agents.