0.541 对 0.651，0.580 对 0.795，28.7 对 45.3。三组数字分别来自 OfficeQA、CL-Bench 和 ServeLearnBench 三个 Agent 基准，等号左边是什么都不装的裸跑 Agent，右边是加了一层外挂之后的同一个 Agent。模型没换，权重没动，连 Agent 的主程序都没改一行，分数却实打实地上去了，ServeLearnBench 那一档甚至把榜单上原有的最佳学习方法甩开将近六成。
做到这件事的东西叫 harness 适配器，把它做成完整开源框架的是 Infini-AI-Lab，一个专注于 Agent 基础设施的研究团队。2026 年 10 月 25 日，他们放出 Hermit v0.1.0，CMU 博士后郑海中两天后在 X 上正式官宣。框架走 Apache-2.0 协议，PyPI 一行命令装完，自身不带任何参数，全部工作都发生在 Agent 的外壳上。这个时间点很微妙：过去一年大家都在卷模型本身，而最近几个月，行业注意力明显开始向模型外面的东西转移，记忆、技能、上下文工程轮番成为热点，Hermit 把这条线推到了新位置——外壳本身也可以是被训练出来的。
先对齐概念：什么是 harness，什么是适配器
Agent 圈里说的 harness，指的是套在模型外面那层执行外壳：循环控制、工具调用、上下文拼装、错误恢复，全在这层。Claude Code 和 Codex 各自有一套 harness，模型的能力正是被这层外壳翻译成实际行动的。同一个模型，换一套 harness，跑出来的任务成功率可以差出一大截，这在今年的多个公开对比里已经被反复验证过。举个具体的例子：模型给出一段有语法错误的补丁，一套外壳直接报错退出，另一套外壳会把编译错误喂回去让模型再修一轮，第三套还会先跑一遍测试再决定要不要重试。三种外壳下，模型的智力没变，任务结局完全不同。外壳里沉淀的其实是工程经验，只是这些经验一直用手写的方式维护。过去想让 Agent 更擅长某类任务，主流做法只有两条路：要么微调模型，比如 LoRA 那套参数高效微调；要么改 harness 源码，往里加技能、塞提示词。前者又贵又慢，一张卡都未必够；后者有两个现实障碍——很多 harness 根本不开源，而开源的那些内部组件调得极细，动一处崩一片，升级上游版本时还要手工合并自己的改动。
Hermit 走的是第三条路：模型和 harness 都冻结，单独训练一个外挂的适配器。这个适配器是带训练支持的插件，里面可以装指令、技能、记忆、工具、可执行代码，甚至多 Agent 的协调逻辑，规格文件里写清楚哪些部分可学习、更新怎么校验。它更像一个完整的仓库，而不是一段提示词。团队给出的设计理由很直白：很多 harness 闭源改不了；开源的改了容易破坏调好的组件协作；而学出来的适配器可以在兼容的模型和 harness 之间搬家，换基座模型时不用重新训练一遍。
训练循环：Trainer、Judge、Evolver 三角
Hermit 的训练是一个强化学习风格的闭环。具体走一轮是这样的：Trainer 拿着当前版本的适配器，把一批任务喂给 Agent 真跑一遍，每条轨迹都完整记录模型看到什么、调了什么工具、在哪一步开始跑偏；Judge 按任务的成功标准给输出打分；Evolver 读完整轨迹和分数，分析失败模式，提出下一版适配器，可能是补一条排错指令，可能是换掉一个技能，也可能是重写整个协调流程；Trainer 再按规格模板校验新版本的合法性，通过后提交为下一版。一轮接一轮，适配器在真实任务反馈里迭代，而不是靠人拍脑袋改提示词。每个适配器版本都挂在一棵 git 树上，可以检视、回滚、发布任意版本，工程属性直接拉满——训练产物不再是看不懂的权重文件，而是可以 diff、可以 code review 的文本资产。学习程序还能放进沙箱里跑，这一步要求 Linux 加 bubblewrap。
框架内置五种进化方法：GEPA、SkillOpt、元世界模型 MWM、Harness² 和 Continual Harness，分别来自五篇不同的研究论文，ACE 与 ReHarness 标注为即将支持。一个框架把五种方法的训练循环统一成同一套接口，这件事本身就省掉了大量对比成本。嫌不够的话，Evolver 接口是开放的，可以自己写一个插进去。这五种方法对应不同的学习策略：GEPA 和 SkillOpt 先用一小批任务训练、然后冻结评测；Continual Harness 和 Harness² 则支持边跑边学，在流式任务里持续更新自己。覆盖面已经超出调提示词的范畴，进入了持续学习的地盘。
自报成绩：涨分是真的，水分也要挤
数字全部来自仓库示例 README，跑分模型统一是 GLM 5.3 Flash，而且大多是单次运行，这个前提先记住。
几个细节值得拆开看。OfficeQA 的 172 道测试题全部围绕美国国债公报展开，用精确匹配判分，考的是 Agent 在陌生文档集合里找答案、管流程的基本功。OfficeQA 上 GEPA 和 SkillOpt 只用 50 道题训练然后冻结测试，分别拿到 0.580 和 0.600，说明即使是冻结式学习也能从轨迹里榨出有效信息；Continual Harness 和 Harness² 则是边流式过测试集边学，Harness² 相对自己的参照运行提升 7.6 个百分点，95% 置信区间落在 +1.7 到 +13.4。CL-Bench 出自今年那篇持续学习基准论文，arXiv 编号 2606.05661，测的是有状态环境下的学习能力，数据库探索和可利用扑克两个子项一个考知识积累、一个考策略博弈，MWM 两边都拿了第一。ServeLearnBench 模拟的是更贴近生产的场景：624 条零售客服请求背后藏着一套会漂移的策略，Agent 必须在服务过程中自己摸出规律。给定隐藏策略的预言机上限是 89.2，MWM 的 45.3 离天花板还远，但已经把榜单上最好的学习方法 CH 的 28.7 甩开一大截。
反面数据同样摆在那里：纯提示词适配器在 CL-Bench 数据库任务上只得 0.470，比不学的基线还低。这说明这层外挂不是随便塞点话就管用，训练方法本身是胜负手，把静态提示词换个地方存放不会产生任何增益，甚至可能因为占用了上下文窗口而帮倒忙。
和同类摆在一起：GEPA、ACE、Agent Lightning
训练对象的差异是根本差异。GEPA 优化的是提示词和代码这类文本参数，ACE 维护的是上下文资产，Agent Lightning 直接碰权重，只有 Hermit 把 Agent 的行为外壳本身当成可学习对象。代价与收益也跟着不同：不碰权重意味着训练便宜、迁移容易，一个适配器可以在兼容的模型和 harness 之间复用；但能力天花板也受基座模型限制，外壳调得再好，模型本身不会的推理它还是不会。参照成绩上，GEPA 在 AIME 2025 上把 GPT-4.1 Mini 从 46.6% 推到 56.6%，ACE 在 AppWorld 上报告了 +10.6% 的提升，Hermit 拿出的则是 CL-Bench 数据库任务的 0.795。三家各有主场，选择时先看自己要改的是文本、上下文、权重还是外壳。
上手：从 pip 到把适配器插进 Codex
安装就是一行，PyPI 包名 hermit-harness，纯 Python 框架，自己一个参数都不带：
pip install hermit-harness
CLI 覆盖 check、train、eval、deploy、convert 五个动作。离线 Quick Start 不需要 GPU、不需要模型账号和数据集，用内置的模拟流程就能把训练闭环走完一遍，适合先摸清楚各阶段产物长什么样；真跑训练时后端任选，官方支持名单里有 Fireworks、OpenAI、Anthropic、DeepSeek，本地则走 vLLM、SGLang 或 Ollama，模型引用统一写成提供方加模型名的形式。任务集就是一个 jsonl 文件，每行一条任务，配上对应的评分标准，把你日常让 Agent 反复干的活整理进去，就能启动第一轮训练。
hermit check
hermit train --method mwm --tasks ./tasks.jsonl
hermit eval --adapter ./adapters/v3
hermit deploy --adapter ./adapters/v3 --target codex
hermit convert --adapter ./adapters/v3 --target claude-code
这五条命令依次对应体检环境、用 MWM 方法训练适配器、评估某个版本、生成 Codex 可直接使用的工作区、把训练好的适配器导出成 Claude Code 原生插件，convert 还支持 Pi 和 Hermes 作为目标。整条链路最顺滑的地方在于产物形态：训练出来的是文本资产，部署出去的是目标工具认得的工作区或插件，中间没有需要自己翻译的格式。仓库里还有个很能说明问题的例子：一个 Lean 协作适配器，指挥 5 个 Codex Agent 合力证明同一条定理，多 Agent 协调逻辑也被整个装进了适配器里。
冷静三问：哪些是真的，哪些还要等
第一，成绩全是团队自报，而且多数是单次运行，没有多种子误差棒，OfficeQA 那条 +1.7 到 +13.4 的置信区间下沿并不算宽，运气好的单次运行和稳定的真实增益之间还有距离。第二，跑分模型只有 GLM 5.3 Flash 这一档，换到更强的基座上适配器还剩多少增益，目前没有数据，基座越强、外挂空间越小的可能性不能排除。第三，v0.1.0 才发版几天，沙箱执行绑定 Linux 加 bubblewrap，Windows 和 macOS 的落地路径还不完整，文档和示例的覆盖面也还在补齐，ACE 和 ReHarness 两个方法还挂在即将支持的名单上。把它当生产组件之前，先在自己的任务分布上跑一遍 eval，是最低成本的排雷方式。
谁该现在动手
如果你在用 Codex 或 Claude Code 跑重复性任务，比如每周都要处理一批相似的代码审查、数据清洗或报告生成，又觉得手写提示词和技能文件维护不动，Hermit 值得马上试：它把这些手工资产变成可训练、可版本化、可回滚的对象，git 树加 CLI 的组合对工程团队非常友好，review 一版适配器和 review 一段代码没有区别。如果你的场景需要权重级别的定制，或者基座模型本身就不够用，那它解决不了问题，Agent Lightning 那条路更合适。把视野放大一点，今年 Agent 工程的主线非常清晰：上下文工程解决了给模型看什么的问题，记忆系统解决了记住什么的问题，而训练循环解决的是怎么让经验自动变厚的问题。Agent 优化的竞争已经从怎么写好提示词，升级到怎么把经验沉淀成可训练的资产：claude-mem 把记忆变成了可训练的资产，Hermit 则把整个行为外壳变成了可训练的资产。这条线上还会长出更多东西，而 harness 适配器是目前形态最完整的开源答案。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
0.541 vs. 0.651 0.580 vs. 0.795, 28.7 vs. 45.3. Three sets of numbers are from OfficeQA, CL-Bench and ServeLearnBench, three Agent benchmarks, with the equal left running naked with nothing on the left, and the same Agent with an extra layer on the right. The model hasn't changed, weight hasn't changed, even Agent's main program hasn't changed a line, but the scores are on the ground, and ServeLearnBench has even thrown out nearly 60 percent of the best learning methods on the list.
The thing that does this is called the Harness adapter, and the complete open source framework is the Infini-AI-Lab, a research team focused on Agent infrastructure. On October 25, 2026, they released Hermit v. 0.1.0, and two days after Dr. Zheng Hai declared it on X. The frame follows the Apache-2.0 protocol, PyPI has completed its line of commands, without any parameters, and all work takes place on Agent 's shell. This is a delicate point in time: everyone has been rolling the model itself over the past year, and in recent months industry attention has clearly begun to shift to something outside the model, memory, skills, context engineering has become a hot spot, and Hermit has pushed the line to a new location — the shell itself can be trained.
The concept of alignment first: what is naturals, what is adaptor
Agent's circle refers to a shell that is placed outside the model: circulatory control, tool call, context assembly, error recovery, all of which is in this layer. Claude Code and Codex each have a set of naturals, and it's this shell that translates into actual action. The same model, a set of harness, could make a big difference to the mission success rate, which has been repeatedly tested in various public comparisons this year. To give a specific example: the model gives a grammatical error patch, a shell is reported as a direct error, and another shell is fed by a translation error Go back to the model for another round, and the third set will run the tests before deciding whether to take retry. Under three husks, the brain of the model remains unchanged, and the end of the mission is completely different. It's actually engineering experience, but it's been maintained by hand. In the past, there were two ways to make Agent better at some kind of task: either fine-tune models, such as LoRA, and fine-tune the parameters; or change the Harness source code, adding skills to the Rig, plugging prompt. The former are expensive and slow, and a card may not be sufficient; the latter have two real obstacles - many naturals are not open at all, while those internal components of the open source are very finely tuned, with a collapse, and the upgrading of the upstream version involves the manual consolidation of their own changes.
Hermit is on the third track: models and naturals are frozen, training an external adapter alone. This adapter is a training-support plugin with instructions, skills, memories, tools, enforceable codes, and even more Agent ' s coordinated logic, and the specifications document contains a clear description of what can be learned, updated and verified. It's more like a complete warehouse than a prompt. The design rationale given by the team is straightforward: a lot of harness is closed; open source changes can undermine well-adjusted component collaboration; and the adaptor learned can move between compatible models and harness without retraining when changing base models.
Training cycle: Trainer, Judge, Evolver Triangle
Hermit's training is a closed circle of enhanced learning style. This is the case: Trainer, with the current version of the adapter, feeds a number of tasks to Agent, runs again, and every track has a complete recording model of what he sees, what tools he transfers, where he starts running; Judge scores the output according to the mission's success criteria; Evolver reads the full track and score, analyzes the failure pattern and proposes the next version of the adaptor, possibly replacing a miscalculation command, possibly changing a skill, or rewriting the entire coordination process; Trainer then validates the new version by the specification template and submits it to the next edition. Round by round, the adaptor is in the feedback of real task, rather than changing prompt by taking a head. Each adapter version is mounted on a guit tree and can be viewed, rolled back, any version of publish, and the engineering attributes are drawn directly - the training product is no longer a weight file that cannot be understood, but a text asset that can diff, code review. The learning program can also run in sandboxes, which requires Linux plus a bubblewrap.
Within this framework, five evolutionary approaches have been put in place: GEPA, SkyllOpt, the meta-world model MWM, Harness2 and Constant Harness, from five different research papers, ACE and ReHarness, each of which are marked as forthcoming. A framework that integrates the training cycles of the five approaches into a single interface has in itself saved considerable comparative costs. If it's not enough, the Evolver interface is open and can be plugged in. These five approaches correspond to different learning strategies: GEPA and SkyllOpt train themselves on a small set of missions and then freeze their assessments; Constant Harness and Harness2 support running and learning, updating themselves continuously in a fluid mission. The coverage went beyond the prompt to the area of continuous learning.
Self-reported: The points are real, the water's gonna squeeze.
The numbers are all taken from the warehouse example README, and the run-down model is GM 5.3 Flash, which is mostly a single run, a prerequisite to remember.
A few details deserve to be taken apart. The 172 test questions for OfficeQA are all based on the US Treasury Bulletin, with precise matching of scores, and Agent's basic work of finding answers and running processes in a collection of strange files. On OfficeQA, GEPA and SkyllOpt are trained only with 50 questions and then frozen with 0.580 and 0.600, respectively, to show that even frozen learning can extract valid information from the trajectory; Continental Harness and Harness2 are overtly tested, with Harness2 increasing 7.6 percentage points relative to their own reference operations, and 95% falling between +1.7 and +13.4. CL-Bench, from this year's study of continuous learning benchmark, arXiv No. 2606.05661, measured learning abilities in a state-of-the-art environment, a database that explores and uses two sub-points of poker to build up knowledge, a test strategy game, and MWM is first on both sides. ServeLearnBench simulates a scenario that is closer to production: 624 retail passenger requests have a drifting strategy hidden behind them, and Agent has to find his own pattern in the service. The upper limit of the predictor for the hidden strategy is 89.2, 45.3 of MWM, far from the ceiling, but the best method of learning on the list has been tossed aside 28.7 of CH.
The reverse data are also there: the pure prompt adapter has 0.470 on the CL-Bench database mission, which is lower than the non-learning baseline. This means that the hang-up does not work simply by talking, that the method of training itself is a win-winner, that the transfer of static prompt to a place would not yield any gain, and that it might even be helpful by occupying context window.
Together with their own: GEPA, ACE, Agent Lightning
Differences in the target audience are fundamental. GEPA optimizes text parameters such as prompt and codes, ACE maintains context assets, Agent Lightning directly touching weights, and Hermit only treats the behavioral shell of Agent itself as a learning object. The costs and benefits are different: not touching weight means training is cheap, moving is easy, and a adapter can be reused between compatible models and naturals; but the power ceiling is also limited by the base model, the shell is better tuned and the model itself does not. By reference, GEOPA pushed GPT-4.1 Mini from 46.6 per cent to 56.6 per cent on AIME 2025, ACE reported an increase of +10.6 per cent on AppWorld and Hermit produced 0.795 for the CL-Bench database mission. Each of the three has its own home, and the choice is between text, context, weight and shell.
Hand in hand: From pip to insert adapter into Codex
Installation is a line, PyPI package name hermit-harness, pure Python frame, without one parameter:
pip install hermit-harness
CLI overwhelms check, track, eval, decloy, convert five actions. offline Quick Start does not need GPU, model accounts and data sets, and a built-in simulation process will complete the training closed loop and will be suitable for first mapping of what the products look like at each stage; the backend is selected for real-time training, and the official support list includes Fireworks, OpenAi, Anthropic, DeepSeek, and the local VLM, SSGLang or Ollama, where the model references are used in the form of a supply-side model name. The job set is a jsonl file, a line-by-line job, with the corresponding scoring criteria, and you can start the first round of training by sorting out what Agent has been doing.
hermit check
hermit train --method mwm --tasks ./tasks.jsonl
hermit eval --adapter ./adapters/v3
hermit deploy --adapter ./adapters/v3 --target codex
hermit convert --adapter ./adapters/v3 --target claude-code
The orders corresponded in turn to the medical examination environment, trained adapters using MWM methods, assessed a version, generated a working area that Codex could use directly, exported trained adapters into Claude Code primary plugins, and supported Pi and Hermes as targets. The smoothest place in the entire chain is in the form of products: text assets are trained, workspaces or plugins identified by target tools are deployed and there are no formats that require translation by themselves. There's also an illustrative example in the warehouse: a Lean Collar Adapter, commanding 5 Codex Agent confluences attests to the same theorem, and many Agent co-ordinated logic is fully integrated into the adaptor.
Calm down and ask: "What's real and what's waiting for?"
First, the results are self-reported by the team, and most are run one-time, without multiple sub-engineers, the OfficeQA +1.7-+13.4 confidence zone is not very wide, and there is still a distance between lucky single-run and stable real gains. Second, only the GLM 5.3 Flash tranche, leaving much to be gained by switching to a stronger adapter on the base, and the possibility that the stronger the base and the smaller the outer space cannot be ruled out without data. Third, v0.01.0 was released only a few days ago, the landing paths for the sandbox enforcement binding Linux plus bubblewrap, Windows and MacOS were incomplete, the coverage of documents and examples was being completed, and the ACE and ReHarness methods were on the list to be supported. Before using it as a production component, run it over the distribution of its tasks, eval, which is the least costly method of demining.
Who should do it now?
If you're running multiple tasks with Codex or Claude Code, such as handling a similar number of code reviews, data cleansing or report generation per week, and feeling that handwritten prompt and skill files are not being maintained, Hermit is worth an immediate test: it turns these hand-written assets into training, versionable, roll-backable objects, Git tree plus CLI combinations are very friendly to the engineering team, review adaptor and review code. If your scene needs to be customized at a weighting level, or the base model itself is not enough, then it won't solve the problem. To zoom in, the main line of this year's Agent project is very clear: the context project solves the problem of looking at models, the memory system solves the problem of remembering, and the training cycle solves the problem of how experience automatically becomes thicker. Agent optimized competition has gone from writing prompt to settling experience into trainingable assets: claude-mem transforms memory into trainingable assets, and Hermit transforms the whole behavioral shell into trainingable assets. There'll be more on this line, and the harness adapter is the most complete open source answer in the current form.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.