[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-google-envharness-agent-environments":3,"news-related-20568e5d-3b66-495d-8b7b-0a702f3c7877":38},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"20568e5d-3b66-495d-8b7b-0a702f3c7877","模型在进化,训练环境却是死的:Google 开源 EnvHarness,给环境也套一层 harness","Google Cloud AI Research 提出 EnvHarness:在静态环境外包一层可编程插件,不改动任务与验证器即可重塑环境;配合黑盒诊断弱点的 EnvRigger 设计器,在 4 领域 5 基准上 held-out 最高 +9.0 分、步数省 9.8%,并支撑模型-环境持续共进化。","过去两年,Agent 领域的大部分创新都发生在模型这一侧:给模型套上工具调用、记忆模块、技能库,一个静态的 LLM 就变成了能自主干活的 agent。但训练环境的另一侧,几乎没人动过——基准环境是一次性手工搭好的,任务固定、反馈固定,对所有 agent 一视同仁,agent 变强之后环境就再也教不了它什么。Google Cloud AI Research 8 月 20 日发布的 EnvHarness(arXiv:2608.19880),把 agent harness 的思路镜像到了环境侧:在冻结的环境外面套一层可编程插件,不改底层逻辑就能重塑环境行为。论文提交当天即登上 Hugging Face Daily Papers 榜首,目前 214 个 upvote,代码已开源(github.com\u002Fgoogle-research\u002Fenvharness)。\n\n## 核心思路:包装,而不是重造\n\nEnvHarness 的关键设计是「wrapping, not authoring」。它通过标准的 reset() \u002F step() 接口操作环境,提供三类可自由组合的插件组件:Stage 定制回合的起始状态,Contract 改写状态转移规则(比如强制「提交前先跑测试」),Chain 修改观测链路。组件之间满足组合律,E″ = w₂(w₁(E)),叠多少层都不破坏接口。最重要的一条约束:每个被重塑的环境都保留原基准的人工验证器——这直接回应了「环境生成需要昂贵或不可靠的验证器」的行业痛点。改造后的环境可信度继承自原基准,不需要为生成环境重新造一套判分系统。\n\n## 自动化设计器:EnvRigger 把策略当黑盒\n\n手动写插件仍然昂贵,论文的 EnvRigger 把这步也自动化了。它把目标策略当黑盒,观察其执行轨迹,诊断出系统性缺陷(例如「agent 提交补丁前从不先跑失败测试,成功靠的是运气而不是验证」),然后针对缺陷合成 EnvHarness 组件,再通过全新 rollout 验证组件是否真的有效。一个完整的设计循环:观察 → 诊断 → 写组件 → 验证 → 保留或修订。验证器自始至终未被触碰。\n\n## 数字账本\n\n在 4 个领域 5 个基准上,EnvHarness 全面超过原始环境和领域专用环境生成管线:held-out 实例最高提升 9.0 分,同时执行步数减少 9.8%。项目页给出了更细的拆解:WebArena 上 38.7 → 41.6,SWE-bench Verified 上 47.7 → 52.6,ALFWorld 上 61.7 → 68.3(均为 base agent → EnvHarness 环境学习)。在 SWE-bench Verified 的环境规模实验里,同样约 50 个环境的规模下 EnvHarness 达到 54.79,超过 SWE-smith 的 50.37 和真实环境的 52.09。跨模型实验覆盖 Gemini 3.1 Flash-Lite、Qwen3.6 27B、Gemini 3.5 Flash、Claude Sonnet 4.6 四个策略模型,增益稳定在 +2.9 到 +3.7 之间,base 跨度 30.7 到 67.2。RL 场景下,在 EnvHarness 环境上做强化学习也优于在原始环境上做,ALFWorld 上 88.3% vs 85.4%;三轮共进化把 SWE-bench Verified 从 47.7 推到 54.8。还有一项很实用的能力:按需控制难度,把基准成功率调进指定区间 [0.4, 0.6],论文报告 6% → 80% 的可控率。\n\n## 我的看法\n\n这篇论文最大的价值不是某个具体数字,而是把「环境侧的可编程性」变成一个正经的研究对象。自进化 agent 方向的工作大多在改提示词、改技能库、改脚手架,而 agent 所处的世界保持冻结——一个不断变强的 agent 对着冻结的基准,迟早学到无物可学。EnvRigger 的黑盒诊断 + 合成插件流程,等价于给训练环境装上了「自动出题人」,并且每道新题都自带人工验证器背书。对做 agent RL 的团队来说,这条路线比从零生成环境便宜得多,也更容易通过合规审查。局限也很明显:数字全部来自论文自报,第三方复现尚未出现;组件类型只发布了三种,接口还能容纳更多;而「按需调难度」如果被滥用到 eval 上,也可能变成新的刷榜工具——环境可控与 eval 可信之间的张力,值得整个社区盯紧。\n\n原文:https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.19880","https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.19880","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"8cf7490f-2449-4ba7-be19-61befa0d92b4","google",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"68bdf799-d0c0-4b66-a2e7-31d16a5fff66","en","Google Open-Sources EnvHarness: Giving Agent Environments a Harness","Google's EnvHarness makes static environments programmable via plug-ins that leave tasks and verifiers intact, gaining up to 9 points on held-out benchmarks.","For the past two years, most agent-layer innovation has happened on the model side: wrap a static LLM with tools, memory, and skill libraries, and it becomes an autonomous agent. The other side of the loop — the training environment — has barely moved. Benchmark environments are hand-built and static: fixed tasks, fixed feedback, identical for every agent, and quickly exhausted once the agent improves. EnvHarness, released on August 20 by Google Cloud AI Research (arXiv:2608.19880), mirrors the agent-harness trick onto the environment side: a programmable layer of plug-in components that wraps a static environment and reshapes its behavior without modifying the underlying logic. The paper hit #1 on Hugging Face Daily Papers on submission day and currently sits at 214 upvotes, with code open-sourced at github.com\u002Fgoogle-research\u002Fenvharness.\n\n## The core idea: wrapping, not authoring\n\nEnvHarness's key design choice is \"wrapping, not authoring.\" It operates through the standard reset() \u002F step() interface and ships three freely composable component types: Stage customizes where an episode starts, Contract rewrites transition rules (for example, forcing \"run the tests before submitting\"), and Chain modifies the observation pipeline. Components compose — E″ = w₂(w₁(E)) — and stacking never breaks the interface. The single most important constraint: every reshaped environment retains the original benchmark's human-built verifier. This directly answers the industry pain point that environment generation depends on expensive or unreliable verifiers. The reshaped environment inherits trust from its source benchmark, with no need to rebuild a judging system for generated environments.\n\n## The designer: EnvRigger treats the policy as a black box\n\nHand-writing plug-ins is still expensive, so the paper's EnvRigger automates that step. It treats the target policy as a black box, observes its execution trajectories, diagnoses a systemic flaw (for example, \"the agent submits patches without running the failing test first — success depends on luck, not verification\"), synthesizes an EnvHarness component targeting that flaw, and validates it via fresh rollouts. The full design loop: observe → diagnose → write → validate → keep or revise. The verifier is never touched.\n\n## The numbers\n\nAcross five benchmarks in four domains, EnvHarness outperforms both original environments and domain-specific environment generation pipelines: up to a 9.0-point improvement on held-out instances, with 9.8% fewer execution steps. The project page breaks it down further: WebArena 38.7 → 41.6, SWE-bench Verified 47.7 → 52.6, ALFWorld 61.7 → 68.3 (all base agent → learned in EnvHarness environments). In the SWE-bench Verified environment-scaling experiment, at a comparable scale of roughly 50 environments, EnvHarness reaches 54.79 versus 50.37 for SWE-smith and 52.09 for real environments. The cross-model experiment covers four policy models — Gemini 3.1 Flash-Lite, Qwen3.6 27B, Gemini 3.5 Flash, and Claude Sonnet 4.6 — with gains holding steady between +2.9 and +3.7 while no-skill baselines span 30.7 to 67.2. Under RL, training on EnvHarness environments beats training on the originals (88.3% vs 85.4% on ALFWorld), and three rounds of policy-environment co-evolution push SWE-bench Verified from 47.7 to 54.8. One more practical capability: difficulty on demand — steering a benchmark's success rate into a requested band [0.4, 0.6], with the paper reporting 6% → 80% controllability.\n\n## My take\n\nThe paper's biggest contribution is not any single number but making \"environment-side programmability\" a first-class research object. Most self-evolving agent work rewrites prompts, grows skill libraries, and edits scaffolds while the world the agent evolves against stays frozen — an improving agent facing a frozen benchmark eventually has nothing left to learn. EnvRigger's black-box diagnose-and-synthesize loop is effectively an automatic problem-setter for training environments, with every new item backed by a human-built verifier. For teams doing agent RL, this route is far cheaper than generating environments from scratch and easier to pass compliance review. The limits are equally clear: every number is self-reported by the paper, with no third-party reproduction yet; only three component types are released while the interface admits more; and if \"difficulty on demand\" gets abused on evals, it could become a new benchmark-gaming tool — the tension between environment controllability and eval trustworthiness deserves the community's attention.\n\nSource: https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.19880","google-envharness-agent-environments","2026-08-20T10:42:06Z","2026-08-21T13:09:13.367809Z","2026-08-21T13:09:13.367820Z",true,"agent",116,{"items":39},[40,45,50,55,60,65],{"id":41,"title":42,"news_slug":43,"published_at":44},"f26ace13-9c96-47ea-a528-b6682a22aa1e","Apodex 1.1 把推理搬进真实执行:PIVOT-RL 定位关键决策点,35B mini 开源","apodex-1-1-agentic-execution-pivot-rl","2026-08-25T14:30:00+00:00",{"id":46,"title":47,"news_slug":48,"published_at":49},"e8965513-b56f-475b-b15f-22a5ea2d2a4e","Agent 取代人成为 HF Hub 一号用户:Claude Code 占 44.4%,还有一次 4.5 天未察觉的入侵","hf-hub-agent-user-claude-code-4-5-day-intrusion","2026-08-21T08:00:00+00:00",{"id":51,"title":52,"news_slug":53,"published_at":54},"b4754043-6b19-499f-8459-f8fc786f4d80","Pokee-Isaac 28B 把 10M 上下文塞进客户边界:28B 参数在 RULER 10M 上 93.3%","pokee-isaac-28b-10m-context","2026-08-20T14:00:00+00:00",{"id":56,"title":57,"news_slug":58,"published_at":59},"36055e5f-136f-497d-8763-3ed6609f59ff","Meta Muse Glimmer 30B 本地落地:Apache 2.0 的开源智能体,把 Agent 装进 24GB 显存","meta-muse-glimmer-30b-local-agent-apache2-r2","2026-08-19T03:00:00+00:00",{"id":61,"title":62,"news_slug":63,"published_at":64},"259d91b2-ed6b-4af8-8f2a-f759b84cc617","蚂蚁 Ling-3.0 Flash：124B\u002F5.1B MoE 的 Agent 生产级模型","inclusionai-ling-3-flash-hybrid-linear-moe-agent","2026-08-14T08:00:00+00:00",{"id":66,"title":67,"news_slug":68,"published_at":69},"5878a668-282c-4b88-b2b8-7eef40b7938c","LFM2.5-2.6B：2.5GB 内存跑本机 Agent 220 tok\u002Fs","lfm2-5-2-6b-on-device-agent","2026-08-11T00:00:00+00:00"]