Adaptive Agent Environments Target Benchmark Saturation

Google's open-source EnvHarness introduces evolving training and evaluation environments through Stage, Contract, Chain, and EnvRigger components, aiming to keep coding-agent challenges difficult as systems improve. Reported gains are promising but remain benchmark-limited, with production transfer, verifier robustness, and anti-gaming properties still unresolved.