谷歌研究人员开源 EnvHarness，这是一个能够根据智能体的短板动态调整智能体训练环境的框架，且无需修改智能体、环境或评估系统。研究人员表示，这种方法既能提高学习效果和效率，又能帮助智能体克服自身缺陷。
研究人员表示，自我进化型智能体通常通过重写提示词和获取新技能来改进自身，但很少会对其运行的环境进行调整。这带来了一个关键问题：
一个在固定基准上不断改进的智能体最终将无法再从该基准中学到东西。对于处于每个学习阶段的每个学习者而言，初始状态、规则和反馈都是相同的，因此一旦智能体能够解决这些任务，环境就再也无法提供新的知识。
研究人员解释说，环境生成成本高昂，需要特定领域的流水线，并且往往需要依赖昂贵或不可靠的验证器。即便投入大量资源，生成出来的环境仍然是静态的。为了让工程师更容易在自适应环境中训练智能体，他们开发了 EnvHarness。
EnvHarness 是一层由可插拔组件构成的可编程封装层，包裹住静态环境，仅改变智能体交互的接口。它通过动态引入变异来重塑环境行为，不需要修改底层环境结构与评估体系。
EnvHarness 使用标准的 reset() 和 step() 接口来包装现有环境，并提供三种可组合组件：setup，用于自定义环境的初始状态；rule，用于根据智能体的动作约束状态转换；以及 link，用于自定义呈现给智能体的观察结果。
该框架还引入了 EnvRigger，它通过分析智能体行为、识别反复出现的薄弱点、生成有针对性的修改方案，并通过新的交互进行验证，自动确定如何针对特定智能体调整环境。
反复运行 EnvRigger 循环能够实现智能体与环境之间持续的共同进化，带来复合式的性能提升，并且该提升效果可随定制任务数量有效扩展。
谷歌的论文提到，添加 EnvHarness 层后，智能体在 WebArena、SWE-bench Verified、ALFWorld 等多个基准测试上的表现均有提升，且无需修改核心逻辑。研究人员测试了多个模型，包括 Gemini 3.1 Flash-Lite、Qwen3.6 27B、Gemini 3.5 Flash 和 Claude Sonnet 4.6，所有这些模型的性能都提升了 2.9 到 3.7 个百分点。
研究人员总结说，EnvHarness 在静态环境无能为力的情况下带来了持续的性能提升，通过领域无关的接口实现了跨领域泛化，并通过修正无效行为提高了效率。
Marketechpost 创始人兼首席执行官 Asif Razzaq 在 LinkedIn 上这样评论 EnvHarness：
它通过标准的 reset/step 接口封装现有环境，原始任务以及人工编写的校验器都可以保持不变。大模型开发者只需要针对智能体运行轨迹中发现的缺陷编写这层封装逻辑。
同样，DevOps 工程师 Aniket Kumar 强调，Stage（setup）、Contract（rule）和 Chain 提供了一种干净的抽象，保留了基准验证器，避免了其他环境扩展方法的一个常见陷阱。
查看英文原文：https://www.infoq.com/news/2026/10/envharness-self-evolving-agents/
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
EnvHarness, an open source for Google researchers, is a framework that can adapt the AI agent training environment to the short board dynamics of AI agent, without the need to modify AI agent, environment or assessment systems. Researchers have indicated that this approach both enhances learning effectiveness and efficiency and helps AI agent overcome its own shortcomings.
Researchers indicated that self-evolutionary AI agent usually improves itself by rewriting prompt and acquiring new skills, but rarely adjusts the environment in which it operates. This raises a key issue:
A AI agent, which is constantly improving on a fixed benchmark, will ultimately no longer be able to learn anything from it. For each learner at every learning stage, the initial state, rules and feedback are the same, so once AI agent is able to address these tasks, the environment can no longer provide new knowledge.
Researchers explained that environmental generation is costly, requires water currents in specific areas and often relies on expensive or unreliable certifiers. Even with significant resources, the resulting environment remains static. In order to make it easier for engineers to train AI agent in an adaptive environment, they developed EnvHarness.
EnvHarness is a programmable containment layer consisting of extractable components, wrapped in a static environment, which changes only the AI agent interactive interface. It reshapes environmental behaviour by dynamically introducing variations, without the need to modify the underlying environmental structure and assessment system.
EnvHarness uses a standard reset() and step() interface to package the existing environment and to provide three combustible components: setup, for the initial state of the custom environment; rule, for action binding based on AI agent; and link, for customizing observations for AI agent.
The framework also introduces EnvRigger, which automatically determines how to adapt the environment to a specific AI agent by analysing AI agent behaviour, identifying recurrent weaknesses, generating targeted modifications and through new interactive validation.
Repeated EnvRigger loops can achieve continuous co-evolution between AI agent and the environment, resulting in complex performance improvements, and the effect of this upgrade can be effectively expanded with the number of custom tasks.
Google's paper mentioned that, with the addition of EnvHarness layers, AI agent had improved performance on a number of benchmark tests such as WebArena, SWE-bench Verifid, ALFWorld and did not need to modify the core logic. Researchers tested several models, including Gemini 3.1 Flash-Lite, Qwen 3.6 27B, Gemini 3.5 Flash and Claude Sonet 4.6, all of which increased their performance by 2.9 to 3.7 percentage points.
The researchers conclude that EnvHarness has brought about continuous performance enhancement in a static environment, with cross-cutting cross-cutting coverage through non-field-related interfaces and efficiency gains through amendments to ineffective behaviour.
Asif Razzaq commented on LinkedIn: EnvHarness:
It covers the existing environment through a standard reset/step interface, and both the original mission and the manual checker remain unchanged. The LLM developer needs only to prepare this layer logic for defects found in the AI agent track.
Similarly, DevOps Engineer Aniket Kumar stressed that Stage (setup), Contractor (rule) and Chain provided a clean abstraction, retained a baseline certifier and avoided a common trap in other environmental extension methods.
See the original English language version at https://www.infoq.com/news/2026/10/envharness-self-evolving-agendas/
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.