EnvHarness Wraps Static Agent Environments in a Plug-In Layer That Rewrites Difficulty Without Touching the Verifier

arXiv 2608.19880·high signalXBlueskyLinkedInCopy link

Google Cloud AI Research posted EnvHarness (arXiv 2608.19880), a programmable layer of plug-in components that reshapes an existing static environment's behavior through standard interfaces while keeping the original verifier intact, plus EnvRigger, which watches a policy's trajectories black-box, synthesizes harness components aimed at diagnosed weaknesses, and validates them with fresh rollouts. Across five benchmarks in four domains it beats both the original environments and domain-specific environment-generation pipelines by up to 9.0 points on held-out instances while using 9.8% fewer execution steps. It topped Hugging Face Daily Papers on 2026-08-21 with 213 upvotes, the clearest sign yet that environment co-evolution, not more hand-built evals, is where agent RL is heading.