EnvHarness: Awakening Static Worlds for Agent Learning

Google Cloud AI Research applies the agent-harness idea to the environment side of agent learning. EnvHarness is a programmable wrapper layer that reshapes a frozen, static benchmark environment purely through its standard reset()/step() interface, so every customized task keeps the original human-built verifier. Three plug-in component types cover three modes of control — Stage (rewrite the initial state to plant obstacles or pre-complete subgoals), Contract (override transition dynamics or observations, e.g. to break repetitive action loops or filter verbose output) and Chain (join environments into longer-horizon tasks). EnvRigger automates the design: it treats the target policy as a black box, observes its trajectories, diagnoses weaknesses, writes candidate components and validates them with fresh rollouts before committing.

Across five benchmarks in four domains (ALFWorld, WebArena, SWE-bench Verified, OfficeQA, SpreadsheetBench), with EnvRigger and the policy sharing the same Gemini 3.1 Flash-Lite / Gemini 3.5 Flash backbone, skills extracted from EnvHarness environments beat skills from the original environments on every benchmark: ALFWorld 62.4 → 68.3 (+9.0 on held-out task types), WebArena 38.5 → 41.6, SWE-bench Verified 49.9 → 52.6 with average steps cut from 53.6 to 49.6 (skills from unmodified environments lengthened them to 55.0). It outperforms the domain-specific generators GenEnv, VeriEnv and SWE-smith wherever they apply, keeps improving as environment count scales where real and generated environments plateau, and yields better GRPO training of Qwen3-8B-base on ALFWorld (85.5 → 88.4) and WebShop. Code released under Apache 2.0.