代理訓練／開源框架
EnvHarness Reshapes Agent Training Environments Without Modifying Verifiers, Raising Out-of-Domain ALFWorld Performance by 9 Points
Teams including Google Cloud AI Research apply the agent harness concept to the environment layer, using composable interfaces to alter initial states, visible information, and task flows. Open-source experiments span five benchmarks and deliver gains of up to 9 percentage points, although reproducing the full results still requires multiple external environments and commercial model APIs.
Teams from Google Cloud AI Research, Washington University in St. Louis, and other institutions have released [EnvHarness](https://arxiv.org/abs/2608.19880), an effort to address the high cost, static content, and rapid model saturation of agent training environments. Rather than regenerating an entire simulator, it adds a programmable wrapper around an existing environment’s `reset`, `step`, observation, and scoring interfaces, while leaving the underlying state transitions and human-written verifiers unchanged.
The [open-source implementation](https://github.com/google-research/envharness) divides components into Setup, Rules, and Link. Setup replays actions to modify the initial state; Rules filter actions, transition results, or observations; and Link inserts a task from another environment into the same episode. The paper uses the names Stage, Contract, and Chain. The concepts are equivalent, but the terminology has already diverged. LLM-generated Rules are executable Python, which the framework runs in isolated subprocesses so that faulty modifications do not terminate an entire batch of experiments.
For automation, EnvRigger treats the agent being trained as a black box. It reads successful and failed trajectories, diagnoses deficiencies, generates environment components, and uses new rollouts to verify that the resulting tasks remain solvable and genuinely target the identified weaknesses. This preserves the original scorer, making results easier to compare than approaches that generate both tasks and verifiers.
The team evaluated the system on ALFWorld, WebArena, SWE-bench Verified, OfficeQA, and SpreadsheetBench. After inducing skills from EnvHarness environments, out-of-domain success on ALFWorld rose from 61.4% with skills learned in the original environment to 70.4%. On SWE-bench Verified, performance increased from 49.88% to 52.58%, while the average number of steps fell from 55.01 to 49.61. Each table reports the mean of three independent runs, but the absolute results were produced using Gemini, and the stored skills also depend on a specific embedding space. Engineering teams should next examine cross-model reproducibility, sandbox boundaries for generated code, and whether policies learned in wrapped environments exploit shortcuts in the interface.