Environment Evolution hardens terminal-agent training

The paper targets a real bottleneck in agent training: once models outgrow static tasks, useful RL signals disappear. Its strongest idea is treating environment difficulty as a controllable training variable rather than endlessly mining model-specific failures.

• –Evolves tasks along scenario novelty, skill rarity, and execution length.
• –Uses verifier-gated multi-agent loops to preserve solvability while increasing difficulty.
• –An Evolution-Lineage Scheduler prevents exposing policies to environments they cannot yet solve.
• –Reported peaks of 71.5% and 64.9% outperform both co-evolution and environment-ensemble baselines.
• –Reproduction remains demanding: the study uses 500 curated seed environments, 200 GRPO steps, and long-context terminal-agent infrastructure.