新智元报道
物理AI正在成为人工智能产业最受关注的方向之一。
而让AI走出屏幕，则需要三样东西。一颗足够可靠的大脑，一套能组织它持续工作的执行系统，以及一个能连接物理设备的身体。模型能力不断进步，Agent执行框架持续完善，但现实世界中形态各异的设备仍然给智能体的理解与调用带来挑战。
围绕模型能力、Agent执行与物理设备连接Agnes AI此次带来了新的技术进展。
Agnes Harness（AGH）正式开源。作为一套面向真实任务的Agent执行基础设施，AGH支持持续任务执行、工具调用、状态管理与结果反馈。同时，AGH体系中的MHS（Model Hardware Standard）设备交互协议及相关组件也同步开放，让开发者能够在同一套执行环境中连接软件工具与可编程硬件。
此外，新一代模型Agnes 3.0 Pro迎来能力升级，其API即将上线。
从模型推理到任务执行，再到物理设备连接，Agnes AI正在将这些能力整合进同一套技术体系，让AI能够参与更复杂的真实任务，并进一步向物理世界延伸。
Agnes AI此次开源的Agnes Harness（AGH），是一套面向真实工作场景的开源Agent Harness，为Agent提供持续运行和多端交付所需的执行底座。
开发者可以通过Web工作台或命令行使用它，也可以接入自己的工具，在此基础上构建面向不同场景的Agent应用。
目前，AGH已在GitHub开放源码及相关开发文档，开发者可以通过源码安装运行，并根据实际需求进一步扩展。
GitHub开源地址：
https://github.com/AgnesAI-Labs/agnes-harness
打开AGH的Web工作台，首先看到的是一个运行在本地环境中的对话界面。开发者可以指定工作目录、发送任务指令，让Agent读取文件、执行命令，再根据执行结果继续修改已有工程。
这一过程依靠AGH的Agent Loop持续推进。模型负责判断下一步操作，Harness则负责调用工具与接收反馈并把结果交回模型，直到任务结束或需要人工介入。
但当任务变得复杂，只看对话记录往往不够。
比如当一个Agent修改了十几个文件，执行过多次命令，最终却没能通过测试时。如何判断是哪一步出了问题？是模型判断失误，还是某个工具执行失败？
在AGH的「轨迹」页面中，开发者可以沿着时间轴回看任务的执行过程。模型调用了什么工具，每个步骤执行了多久，都可以在界面中查看。展开具体记录，还能进一步检查工具参数与执行状态，帮助开发者定位问题。
在扩展能力方面，AGH开放了插件扩展机制。用户可以通过插件接入自己的业务工具，利用MCP连接已有服务，也可以针对不同岗位开发专属界面。
在插件管理页面中，用户能够浏览可发现的插件，查看来源、版本及兼容信息，并进行安装和管理。已经验证过的任务方法还可以通过Skills保留下来，并在其他项目中继续使用。
这些能力也为企业场景的定制提供了基础。
通过FDE（Forward Deployed Engineering）交付方式，工程团队可以深入具体业务现场，根据企业已有系统完成接入与适配。项目中形成的工具、任务方法和业务规则，可以逐步沉淀为可复用的能力，减少后续部署中的重复开发。
AGH还通过App Server让不同使用界面共享同一套后台任务。开发者可以在Web工作台中发起任务，再通过命令行恢复同一会话，不必重新整理此前的操作记录。
从持续执行到业务扩展，AGH为开发者提供了一套可深度定制的Agent运行环境。而当这些工具进一步延伸到机器人、实验仪器等物理设备，Agent的工作范围也随之扩展到现实世界。
除了软件工具与企业系统，AGH还通过MHS将物理设备接入纳入同一套任务执行体系。AGH也进一步从单纯的Agent开发框架，升级为面向多模态大模型、智能体与可编程硬件的开放式基础设施平台。基于MHS支持多类设备接入，让大模型与智能体能够连接真实的物理设备。
Agnes AI团队给AGH布置了一项任务，让办公室里的移动小车找到指定书籍，并通过语音读出书中的内容。
在演示中，小车根据任务要求完成了找书与语音朗读。从接收指令到实际操作，Agent的执行范围开始从软件工具延伸到现实设备。
这背后涉及另一项同步推进开源的技术：MHS（Model Hardware Standard）。
MHS是Agnes AI自主设计的统一设备交互协议，面向各类具有可编程接口的硬件。 通过统一的交互方式，设备可以向Agent提供自身的功能与操作信息，让智能体在面对从未使用过的设备时，也有机会理解其能力，并根据任务要求调用相应操作。
在这一体系中，AGH负责组织整个任务，MHS负责连接具体设备。模型根据目标判断下一步需要做什么，AGH协调工具调用与执行反馈，再通过设备接口完成相应操作。涉及关键步骤时，也可以根据任务要求引入人工确认。
Agnes AI还构建了一座名为Mars World的虚拟火星基地，并在其中设计了一项救援任务：一名宇航员在基地外失去联系，Agent需要利用基地内的设备寻找宇航员，并组织救援。
这是一项基于虚拟设备的模拟实验，重点展示AGH如何组织多设备任务，以及MHS如何为不同设备提供统一的交互方式。在这一模拟场景中，Agent可以调度飞行器搜索目标区域，安排机器人执行救援，并协调基地气闸等设施。整个过程需要结合地图、设备状态与任务反馈，持续推进多个操作步骤。
此次开源的还包括配套的设备接入与开发工具。Agnes AI同步开放了MHS 1.0设备协议、MOS 1.0感知数据协议、AgnesHub设备连接组件、Python与TypeScript开发库，以及mhs-check一致性测试套件。
其中，MHS负责设备状态、工具调用与动作请求，MOS负责相机、传感器、地图等观测数据的传输。开发者可以通过公开的设备开发库接入自己的硬件，并使用一致性测试套件检查协议实现情况。
目前，官方仓库已经提供包含11台虚拟设备的Mars World示例，以及移动机器人、机械臂、传感器等本地模拟设备。开发者无需准备实体硬件，也能体验完整的设备接入与任务调用流程。
从办公室里的移动小车，到虚拟火星基地中的多设备救援，AGH与MHS展示了从真实设备连接到复杂任务协同的不同应用方式。随着更多设备适配得到开发与验证，这套开源体系也为机器人巡检、科研实验与现场作业等场景提供了进一步探索的基础。
MHS开源代码：
https://github.com/AgnesAI-Labs/agnes-harness/tree/main/packages/mhs
MHS设备接入指南：
https://github.com/AgnesAI-Labs/agnes-harness/blob/main/docs/guide/mhs.md
无论是在软件环境中处理复杂任务，还是通过接口与现实设备交互，Agent都需要模型来处理信息并作出判断。
在执行系统与设备连接之外，Agnes AI也公布了新一代模型Agnes 3.0 Pro的最新能力进展，其API即将上线。
在涵盖科学推理、科研编程、指令遵循、业务任务执行及综合知识等能力的六项公开基准测试中，Agnes 3.0 Pro展现出较为全面的能力表现。其中，GPQA取得91.11分，SciCode取得56.13分，与GPT-5.6 Sol（Max）的成绩相差不到1分；在Tau3-Banking模拟银行业务评测中，以48.0分位列本次对比模型第一。
在高难度科学知识与推理评测GPQA中，Agnes 3.0 Pro取得91.11分。GPQA主要考察模型对专业科学问题的理解与推理能力，为衡量模型处理复杂知识任务的表现提供参考。
在科研编程基准SciCode中，Agnes 3.0 Pro取得56.13分，位列本次对比模型前三。SciCode围绕科学研究中的实际编程问题设计，要求模型结合专业知识、推理与代码实现完成任务。在本次对比中，Agnes 3.0 Pro与GPT-5.6 Sol（Max）的57.1分仅相差0.97分，体现了其在科研编程任务中的竞争力。
不同评测反映的是模型在特定任务下的能力表现。对于Agent而言，更重要的是这些能力能否在实际任务中得到体现。
为此，我们还让Agnes 3.0 Pro尝试了一项模拟企业任务：将历史订单服务迁移至新版API，并完成100笔订单的资金日结核验。
除了评测成绩，我们还让Agnes 3.0 Pro尝试了一项模拟企业任务，即将历史订单服务迁移至新版API，并完成100笔订单的资金日结核验。测试材料中预设了重复记录和财务记录缺失等问题。模型需要理解接口文档修改代码，并逐笔核对各项数据。
最终，Agnes 3.0 Pro识别出3笔金额差异、2笔财务记录缺失和4组未决退款。在独立验收中，交付代码通过了10项测试，100笔订单的逐笔结果与预设标准答案一致。对于证据不足的退款，模型保留了未决状态，没有自行选择金额或猜配订单。
在创意开发任务中，Agnes 3.0 Pro制作了一座可交互的热带植物数字温室。从植物的资料整理，到多张植物插画的生成与组织，再到网页布局和交互开发，任务涉及多个连续步骤。最终呈现的页面支持旋转浏览植物标本架、查看单株植物资料，并通过交互探索不同植物的形态特征。
此外，Agnes 3.0 Pro还将一篇天文学论文转化为交互式科普网页。模型重新组织论文内容，并通过行星凌星模拟等交互设计，将专业研究中的复杂概念转化为更加直观的阅读体验。
从复杂业务数据处理到交互应用开发，这些案例展示了Agnes 3.0 Pro在不同类型任务中的实际表现。
Agnes 3.0 Pro在提升模型能力的同时，依然延续了具有竞争力的定价策略。 更低的调用价格，有助于开发者控制长期运行和反复调试的成本，也让复杂AI应用的开发与部署拥有更多尝试空间。
Agnes 3.0 Pro API即将正式开放，参考定价如下：
与此同时，Agnes 3.0 Flash继续保持免费使用。对于需要频繁调试Agent、验证工作流或开展多轮开发实验的用户而言，免费模型能够进一步降低试错门槛，让开发者不必因为Token消耗而限制尝试次数和迭代空间。
从具备复杂任务处理能力的3.0 Pro，到持续免费开放的3.0 Flash，Agnes AI为不同需求的开发者提供了更灵活的模型选择。模型能力与使用成本之间的平衡，也将影响AI应用能否从一次性演示走向持续运行。
3.0 Pro提供模型能力，AGH负责组织任务执行，MHS则将智能体的能力进一步延伸至物理世界。
从Coding、Research到企业系统与真实设备，三者分别面向模型推理、持续执行和硬件连接，为更多Agent应用提供了技术基础。
模型仍在进步，但Agent时代的竞争已经不只发生在模型能力本身。如何将模型、工具、数据、权限、应用和物理设备组织成可靠的系统，正在成为下一代智能基础设施的重要命题。
大脑、记忆和身体，现在都在你手里了。
开源与体验入口：
Agnes Harness：
AGH技术报告：
https://github.com/AgnesAI-Labs/agnes-harness-tech-report
不喜欢
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。
New Ji-won.
Physical AI is becoming one of the most important directions for the AI industry.
And getting AI out of the screen takes three things. A sufficiently reliable brain, an implementation system that organizes its continuous work and a body that connects physical equipment. Modelling capacity continues to improve and the implementation framework for Agent continues to improve, but equipment of varying shapes in the real world continues to challenge the understanding and mobilization of AI agent.
This time, new technological advances were brought about by the connection of Agnes AI to physical equipment around modelling capacity, implementation by Agent.
Agnes Harness (AGH) official open source. As a set of Agent implementation infrastructure for real task, AGH supports ongoing mission implementation, tool call, state management and results feedback. At the same time, the MHS (Model Hardware Standard) equipment interactive protocols and related components in the AGH system are also open simultaneously to enable developers to connect software tools to programmable hardware in the same implementation environment.
In addition, the new generation model Agnès 3.0 Pro, whose API is about to go online, is being upgraded.
From model reasoning to mission execution, to physical equipment connectivity, Agnes AI is integrating these capabilities into a single technical system that allows AI to participate in the more complex real task and further extend to the physical world.
Agnes AI ' s Open Source (AGH) this time is the Open Source Agent Harness for real work scenes, providing Agent with the executive base required for continuous operation and multi-end delivery.
The developers can use it through the Web workstation or command line, or can access their own tools to construct the Agent application for different scenarios.
At present, AGH is in GitHub open source code and related development documents, which can be installed by developers and further expanded according to actual needs.
GitHub open source address:
https://github.com/AgnesAI-Labs/agnes-harness
Opens the web desk of AGH, first seeing a dialogue interface running in a local setting. The developer can specify working directory, send mission commands, give Agent access to documents, execute orders, and continue to modify existing projects based on the results of the execution.
This process has been continuously advanced by AGH's Agent Loop. The model is responsible for judging next step operations, while Harness is responsible for calling on tools and receiving feedback and returning the results to the model until the end of the mission or when manual intervention is required.
When mandates become complex, however, the record of dialogue is often insufficient.
For example, when an Agent modified more than a dozen documents, carried out multiple orders and ultimately failed to pass the test. How can we judge which step went wrong? Was it a model error or a tool that failed?
On the AGH “Track” page, developers can look back at the execution of the mission along the time axis. What tools were used by the model and how long each step was implemented could be viewed in the interface. The development of specific records will also allow for further examination of tool parameters and implementation status and help developer positioning issues.
In terms of capacity expansion, AGH has opened up the plugin extension mechanism. Users can access their own business tools through plugins, use MCP to connect to existing services, and develop a dedicated interface for different jobs.
In the Plugin Management page, users are able to browse to detectable plugins, view sources, versions and compatible information, and install and manage them. The validated mission methodology could also be retained through Skills and continued in other projects.
These capabilities also provide the basis for customizing the business landscape.
Through FDE delivery, the engineering team is able to reach out to specific operational sites and complete access and suitability according to the enterprise ' s existing systems. The tools, mission methodology and operational rules developed in the project can gradually be reduced to reusable capabilities and reduce duplication in subsequent deployments.
AGH also shares a set of back-office tasks with different user interfaces through App Server. The developer can start a mission at the Web workstation and resume the same session through command line without having to reset previous operating records.
From ongoing implementation to business expansion, AGH provides developers with an in-depth customized set of Agent operating environments. As these tools further extend to physical equipment such as robotics, experimental instruments, and so does the work of Agent.
In addition to software tools and enterprise systems, AGH integrated physical equipment access into the same mission performance system through MHS. AGH was further upgraded from a purely Agent development framework to an open infrastructure platform for multimodular LLM, AI agent and programmable hardware. Based on MHS support for multiple types of equipment access, LLM and AI agent were able to connect to real physical equipment.
The Agnes AI team assigned AGH a mission to locate the designated books in mobile cars in the office and to read out the contents of the books by voice.
In the demonstration, the car completed the search for books and voice readings as required by the mission. From the time of receipt of the command to the time of operation, the execution of Agent began from software tools to realistic equipment.
This involves another synchronous propulsion open source technology: MHS (Model Hardware Standard).
MHS is the unified equipment interactive protocol designed autonomously by Agnes AI for various types of hardware with programmable interfaces. By means of a uniform interactive approach, the equipment provides Agent with functional and operational information, giving AI agent an opportunity to understand its capabilities in the face of equipment that has never been used, and to call the corresponding operation as required by the mission.
In this system, AGH is responsible for organizing the entire mission and MHz for connecting specific equipment. The model judged what next step needed to do on the basis of the target, and AGH coordinated tool call with the implementation feedback and then completed the corresponding operation through the equipment interface. Where key steps are involved, human confirmation can also be introduced as required by the mandate.
Agnes AI also built a virtual Mars base called Mars World, in which it designed a rescue mission: An astronaut lost contact outside the base, and Agent needed to use equipment inside the base to find astronauts and organize rescues.
This is a simulation experiment based on virtual equipment, which highlights how AGH organizes multi-equipment missions and how MHS provides a uniform interaction for different equipment. In this simulation scenario, Agent was able to dispatch the vehicle to search the target area, arrange rescue operations for robots and coordinate facilities such as base airlocks. The process will require continuous multiple operational steps combined with maps, equipment status and mission feedback.
The open source also included equipment access and development tools. Agnes AI synchronized the MHS 1.0 device protocol, the MOS 1.0 sensor data protocol, the AgnesHub device connection component, the Python and TypeScript developer, and the mhs-check conformity test package.
Of these, MHS is responsible for equipment status, tool call and action requests, and MOS is responsible for the transmission of observational data such as cameras, sensors, maps, etc. Developers can access their hardware through an open equipment development bank and use a consistency test package to check the protocols.
Currently, official warehouses have provided examples of Mars World, which contains 11 virtual devices, as well as local analogue devices such as mobile robots, mechanical arms, sensors, etc. Developers do not have to prepare physical hardware to experience complete equipment access and mission call processes.
From mobile cars in offices to multi-equipment rescues in virtual Mars bases, AGH and MHS demonstrated different applications that link real equipment to complex mission synergies. This open source system also provides the basis for further exploration of robotics, scientific experiments and field operations as more equipment is developed and validated.
MHS open source:
https://github.com/AgnesAI-Labs/agnes-harness/tree/main/packages/mhs
MHS Device Access Guide:
https://github.com/AgnesAI-Labs/agnes-harness/blob/main/docs/guide/mhs.md
Whether it handles complex tasks in a software environment or interacts with real-life equipment through interfaces, Agent needs models to process information and make judgements.
In addition to the implementation system and equipment connection, Agnes AI has also published the latest capability development of the new generation model Agnes 3.0 Pro, whose API is on line.
In the six open benchmarking tests covering the capabilities of scientific reasoning, scientific programming, compliance with directives, operational task execution and integration of knowledge, Agnes 3.0 Pro demonstrated a more comprehensive performance of capabilities. Of these, GPQA achieved 91.11 points, and SciCode achieved 56.13 points, which is less than one point different from GPT-5.6 Sol (Max); in the tau3-Banking simulation of banking, 48.0 points were ranked first in this comparison model.
In the difficult scientific knowledge and reasoning assessment of GPQA, Agnes 3.0 Pro scored 91.11. GPQA primarily explores the model ' s understanding and reasoning of professional scientific issues to inform the performance of the model ' s complex knowledge management tasks.
In the scientific programming benchmark SciCode, Agnes 3.0 Pro scored 56.13 points, ranked in the top three of the comparison model. SciCode was designed around practical programming issues in scientific research, requiring models to combine expertise, reasoning and codes to accomplish their tasks. In this comparison, the 57.1 points of Agnes 3.0 Pro and GPT-5.6 Sol (Max) are only 0.97, reflecting their competitiveness in scientific programming.
Different assessments reflect the performance of the model ' s capabilities under a given mandate. For Agent, it is more important that these capabilities be reflected in the actual mandate.
To that end, we have also asked Agnes 3.0 Pro to experiment with a simulation business mission: to relocate historical order services to the new version of API and to complete a fund-day nodule survey of 100 orders.
In addition to evaluating results, we have also asked Agnes 3.0 Pro to experiment with a simulation business mission to relocate historical order services to the new version of API and to complete a fund-day nodule survey of 100 orders. Issues such as duplicate and missing financial records were pre-setd in the test material. Models need to understand the interface document to change the code and check the data one by one.
Ultimately, Agnes 3.0 Pro identified three discrepancies, two missing financial records and four outstanding refunds. In stand-alone acceptance check, the delivery code passed 10 tests and 100 order-by-order results were consistent with preset standard answers. For refunds for insufficient evidence, the model retained the outstanding status and did not select the amount or guess the purchase order.
In the creative development mission, Agnes 3.0 Pro produced an interactive digital greenhouse for tropical plants. From plant documentation to the generation and organization of several plant illustrations, to web layout and interactive development, the tasks involve several successive steps. The final page supports the rotational browsing of plant specimen frames, the viewing of single plant data and the interactive exploration of morphological characteristics of different plants.
In addition, Agnes 3.0 Pro converted an astronomical paper into an interactive science web page. The model reorganizes the content of the paper and transforms the complex concepts of professional research into a more intuitive reading experience through interactive design, such as planetal angglomerations.
From complex operational data processing to interactive application development, these cases demonstrate the actual performance of Agnes 3.0 Pro in different types of missions.
Agnes 3.0 Pro continues competitive pricing strategies while upgrading model capabilities. Lower call prices help developers control the costs of long-term operation and repeated debugging, and give more trial space for the development and deployment of complex AI applications.
Agnes 3.0 Pro API will soon be officially opened with reference pricing as follows:
Meanwhile, Agnes 3.0 Flash continues to be used free of charge. For users requiring frequent debugging of Agent, validation of workflow or multiple rounds of development experiments, free models can further lower the test error threshold so that developers do not have to limit the number of attempts and iterative space because of Token consumption.
From 3.0 Pro, with complex task handling capabilities, to the 3.0 Flash, Agnes AI, which continues to be open free of charge, provides more flexible model options for developers of different needs. The balance between model capabilities and usage costs will also affect the ability of AI applications to move from one-time demonstration to continuous operation.
3.0 Pro provides modelling capability, AGH is responsible for organizing mission execution, and MHS further extends the capabilities of AI agent to the physical world.
From Coding, Research to Enterprise Systems and Real Equipment, they are oriented towards model reasoning, continuous execution and hardware connectivity, respectively, providing a technical basis for more Agent applications.
Models are still progressing, but Agent's competition has gone beyond model capabilities themselves. How to organize models, tools, data, permission, applications and physical equipment into reliable systems is becoming an important proposition for the next generation of intelligent infrastructure.
The brain, the memory and the body are in your hands now.
Open source and access:
Agnes Harness：
AGH Technical Report:
https://github.com/AgnesAI-Labs/agnes-harness-tech-report
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.