最近，Personal Agent「千问Go！」掀起了一阵热潮，参与测试的用户经历被反复转发：用户只说了一句“帮我沟通赔偿”，它就去查订单、核对航班动态、联系客服，甚至把平台原本只肯给的代金券，争取成了70元现金。而这并非孤例：市场上也正推出同类产品，Meta的Muse、Manus的Cue，以及国内多家厂商的产品密集面世。
这一系列信号正在印证，Agent已经从走出演示阶段，走进日常事务，规模化落地已经开始。今年5月，Gartner预测，AI智能体软件支出将从2025年的864亿美元，增至2026年的2065亿美元，并于2027年达到3763亿美元。
但要实现规模化，需要满足一个更为现实的前提：上一个技术时代，云原生成为了应用运行的标准，也定义了云的技术形态。如今，Agent开始规模化落地，需要的是“另一朵云”。
区别在于任务的形态：一个回答问题的应用，可以在返回文本后结束请求；一个替用户办事的Agent，要等待外部系统响应，在中途生成文件，再根据反馈调整下一步，用户关掉界面，任务也未必结束。当这样的任务以亿为单位运行，云要交付的就不再只是算力和模型，而是一套能让Agent持续运行的环境。
国际领先的云厂商已经开始集体换道。研究机构Omdia在《Agentic AI Cloud Titans in Asia &Oceania,2026》报告中，把AWS、阿里云、Google Cloud和微软Azure评为Agentic AI云的四家领导者，他们几乎在同一时间完成判断，并开始调整自己的技术栈。Google Cloud在Next大会上宣布从Cloud Native转向Agent Native，AWS推出Bedrock AgentCore，微软Azure的Foundry Agent Service正式商用，阿里云更是明确提出Agentic Cloud。
以阿里云为例，自今年5月提出向Agentic Cloud转型之后，其技术栈就开始快速调整和迭代，短短四个月后的云栖大会上，阿里巴巴CEO吴泳铭把方向说得更直接，“芯片负责生产Token，云负责把Token送到每一个需要它的地方”，要建设的是“为模型和Agent持续运行而设计的AI云”。
在他的判断中，这朵云还会形成网络效应，计算资源与工具得到联合优化之后，更多用户和Agent会加入，进一步扩大网络。
这提供了一个类似梅特卡夫定律的观察角度，该定律用网络价值与用户数量平方之间的关系，描述连接可能产生的价值。Agent能使用的工具越多，能够获得的业务信息越充分，彼此协作越顺畅，它可以承担的任务就会倍增，创造的经济价值也越大。
同一场大会上，阿里云CTO李飞飞进一步解释了这套体系如何建设，“一个是生产智能，一个是创造价值”，前者决定模型能力如何被生产和供给，后者回答这些能力怎样在企业里用起来。而两者之间，就是Agentic Cloud需要交付的内容。
Agent规模化落地，云要交付的不再只是资源
Agent基础设施会成为云厂商共同投入的领域，是因为它交付的不是一个调用一次就结束的平台接口，而是一段能跑到结束的运行过程。
这类任务的数量还在增长。Gartner预测，到2028年，33%的企业软件应用将内嵌Agentic AI，2024年这一比例不足1%，届时至少15%的日常工作决策将由智能体自主完成。
市场上各家给出的路径并不相同。一些超大云厂商把智能体的运行环境与企业级编排做成了云上的标准产品，另一些厂商则从沙箱、运行时、Token供给这些单点切入。方向是一致的，都在把云改造成能承载Agent的基础设施。
对企业来说，用不用这些现成的能力，取决于自己搭建的代价。自行搭建，就要连接模型和业务系统，管理状态与权限，处理异常，并为每次改动建立评测机制，应用越多，维护这些连接的工作越重。云厂商提供的基础能力，能够替企业承担这部分工作，企业仍然需要定义业务规则和验收结果，但不必为每个Agent重新建设运行环境。
阿里云的选择是做Agentic Cloud，把这些共性工程一起纳入云的交付范围。
这一决策源于对用户需求的洞察，既要考虑模型如何被生产，也要考虑企业如何构建和使用智能体。李飞飞表示，模型能力决定智能决策水平，Context提供决策依据，Harness为执行和持续迭代提供保障。
但要实现文章开头吴泳铭提到的AI云的网络价值，前提是单点足够可靠。一个模型作出错误判断，可能是推理能力不足，也可能是拿到了旧数据；一个计划没有执行成功，原因则可能出在工具调用或运行环境。只提高模型分数，并不能自动解决其他环节的问题。
阿里云因此把能力建设延伸到模型训练与推理，智能体的构建与运行，以及企业数据的组织和使用，相比采购一项产品，企业得到的是一套可以共同工作的基础设施。
Agentic Cloud如何接住
企业交给Agent的是一项完整任务，能不能把活儿干好，取决于模型要判断怎么做，上下文要提供准确的业务信息，Harness则要组织工具调用，并根据执行反馈调整步骤。任何一环跟不上，都可能让任务停在半路。Agentic Cloud要解决的，就是让这三个要素在构建与运行中配合起来，并支持后续的治理和迭代。
李飞飞在云栖大会上进一步拆解了这个问题，结合客户反馈和落地实践，他将企业级智能体需要的支撑归纳为四大能力支柱。
第一项是全域认知引擎，解决Agent依据什么作出判断的问题。
Agent的判断质量，很大程度取决于它能看到什么数据，而不只是模型有多强。
企业通常有大量数据，却未必能直接用于任务，资料分布在不同系统里，同一个概念可能有不同表达，业务状态也在不断变化。模型需要知道哪些信息有效，哪些内容与当前任务有关，以及它是否有权使用这些信息。
阿里云通过数据库和大数据能力处理这类需求，OpenLake提供全域多模态数据湖仓，解除引擎和数据的绑定，支持数据和AI引擎可控平权消费同一份数据。 ApsaraLakebase以湖库一体、数智融合的架构，将数据湖与数据库能力结合，并支持数据版本管理。Agent Context的上下文工程进一步把散在各处的资料整理成任务需要的信息，能被Agent直接使用。
数据还要持续更新，一次人工修正，一条新的业务规则，都可能改变下一次任务的判断依据，企业需要管理的因而包括原始资料，也包括执行过程中积累的经验。
第二项是自主运行与持续进化，对应Harness的主要职责。
Agent运行不是一次调用就能完成的，中间要跨多个工具、多轮执行。Harness提供的正是执行框架，通过组织模型进行工具调用、安排执行步骤，再根据结果调整下一步。
阿里云构建了面向企业级智能体的构建与运行平台AgentCore，把Agent需要的基础设施标准化、托管化，支持长任务、失败重试、断点恢复和异步执行，主要能力包括：执行环境中的沙箱负责隔离，存储保存状态与文件，监控则帮助团队了解任务经过了哪些步骤。
这些能力解决的是不同问题，但需要共同工作：任务中断后，如果文件和状态无法恢复，Agent就可能重新执行已经完成的操作；调用失败后，如果缺少记录，团队也很难判断应该修改模型提示，还是调整工具与环境。
第三项是可信协作与治理，对应权限层。
Agent用的是企业的系统和数据，企业要先弄清楚它是谁、能碰什么，才敢把业务真正交给它。
Agent进入企业系统后，需要明确身份及权限，它代表谁执行任务，哪些操作允许自主完成，哪些必须经过确认，都要有具体规则，多个Agent共同工作，还会涉及信息如何共享，以及责任如何追溯。
阿里云的智能体身份服务Agent Identity负责为身份与授权提供支撑，结合安全机制约束Agent的实际操作，可观测产品则帮助企业分析调用过程和异常，它们共同决定企业能够在多大范围内放手使用Agent。
李飞飞表示，“信用、信任以及支付能力非常关键，我觉得这一层的重要性将会越来越凸显。”一个任务是否被正确授权，调用了什么工具，产生了什么结果，都应当可以检查。
第四项是智能实时决策，解决的是让前三项能够真正接入业务现场。
当业务信息能够及时进入上下文，模型具备判断能力，运行与治理机制也准备就绪，Agent才能在企业工作流中根据实际状态采取行动。
要实现上述系统化运转，就要考验底层能力：模型响应要快，推理成本要低，Agent才能在业务节奏里反复调用。阿里云通过灵骏支撑模型训练和推理，CPFS处理大规模训练中的数据读写——训练要保存检查点，故障后再恢复，读写速度决定了昂贵的计算资源的利用率，推理环节则需要处理资源调度和长上下文等问题，如果请求在队列里等太久，前面的判断再准也用不上。
上述四项能力分属不同层面，李飞飞也强调，沙箱和存储提供执行条件，Harness组织任务及反馈，企业数据形成上下文，如果将它们全部归入某一个工具，反而无法解释实际工程中需要解决的不同问题。
工程落地才是试金石
能力体系是否成立，要看它在真实业务里的表现，而真实业务考验的是工程。
模型给出的判断是概率性的，同一个模型接进不同的系统，结果差别可以很大；企业要的是任务被稳定完成，从数据的组织、任务的执行，到出问题后的定位与修复，每个环节都要靠工程补齐。
三家企业，三种工程难题，比架构图更容易说明，这些能力为什么需要配合着使用。
稀宇科技（MiniMax）与阿里云的合作覆盖了从模型训练到推理应用落地的完整链路。
强化学习需要让模型自己进入环境改代码、编译、跑测试——失败了基于反馈修改，成功了保存完整轨迹进入下一轮。数据由此从静态变为动态，对沙箱的需求远超传统云计算：千万级沙箱快速加载、启动销毁压缩到几秒之内、应对一分钟从十万级到百万级的弹性波动。阿里云Agent Sandbox可以每分钟十万级并发拉起、支撑百万级并发运行，启动约一到两秒，热加载在一百毫秒以内。
判断模型是否真的变强了，同样依赖基础设施。评测要让模型进入真实环境完成复杂任务，每个检查点都要拉起大量沙箱反复采样；计算和存储不稳定，就分不清失败来自沙箱抖动还是模型能力。MiniMax研发工程师、开发者关系负责人李元说，弹性控制成本之后，就能以足够低的成本知道模型到底是变强了，还是只是分数变高了。
模型能力还要变成稳定的产品。Agent和聊天机器人最本质的区别，是有没有执行环境和状态——用户写了一半的代码、中间的文件和上下文，都需要完整保存。MiniMax此前采用对象存储方案，随着用户文件量增加，等待时间增长到十几秒，变得不可接受；向阿里云反馈需求几个月后，AgenticFS发布，每个用户拥有独立的性能配额，可直接挂载Agent Sandbox。李元说，最有价值的不是买云产品，而是云基础设施可以和团队一起快速迭代。
小鹏集团则是与阿里云共建企业数据底座，把分散的业务知识组织成Agent可用的上下文。
小鹏集团AI前瞻专家伦致远介绍，随着Agent进入实际工作流程，团队需要处理不断增长的知识，支持不同Agent的数据空间，并让记忆覆盖更多业务信息。这个数据底座覆盖业务知识和训练数据等需求，ApsaraLakebase是其中的支撑之一。业务规则和历史资料经过组织后，Agent才可能准确理解请求并完成操作。
据伦致远现场披露，小鹏集团相关实践中的缺陷发现到修复耗时，由两天缩短至十分钟，在设计场景中，团队把历史素材和评审过程，以及设计师的理念提供给Agent，单次生成合规率达到83.3%。
海尔盈康一生与阿里云通过可观测体系建设和Harness优化，解决了任务定位和成本匹配的问题。
海尔盈康一生CAIO倪永全介绍，企业级办公助手小盈已经服务约13000名员工。应用规模扩大后，团队需要知道任务在哪一步出现问题，也要避免为简单请求使用成本过高的模型。为此，海尔盈康一生基于阿里云AI网关和AgentLoop建设了可观测与分析体系。在每一次推理可观测的基础上，构建了“观测、评估、优化”的Harness工程闭环，并根据用户意图选择模型，使计算投入与任务难度相匹配。目前，相关机制使任务成功率提升10个百分点，平均模型调用成本下降超过30%。
AgentCore被用作运行底座，配合沙箱隔离、安全检测与授权制来约束实际操作。这种效果的提升来自多个环节的系统性共同调整，而非某一项产品的单一提升。
此外，海尔盈康一生还把内部应用的业务能力提供为Agent可调用的工具，把流程经验沉淀为技能，应用产生的数据则进入企业数据湖，供其他工作继续复用。这种能力的复用，与AI云的网络效应直接相关。一项业务能力建设完成后，如果能够被更多任务调用，其价值就不局限于最初的应用。但连接必须有用，信息共享也需要符合权限边界；简单增加Agent数量，无法保证业务收益的同步增长。
三家企业分别把问题落到了模型运行环境的工程供给、企业上下文建设和Harness工程闭环，它们需要解决的细节不同，共同之处是持续投入工程，才能让模型能力稳定地用于业务。
Agentic Cloud的经济价值就体现在这里，企业的目的是让工作以更可靠、更经济的方式完成，阿里云已经通过产品体系和工程实践，将模型、Harness与上下文的协同落实到真实业务。
对企业而言，当交付周期缩短了，任务成功率提高了，成本下降了，Token才在企业的账上有了明确的价值。（本文首发于钛媒体APP）
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。
Recently, Personal Agent’s “Go!” has generated a surge, and the user’s experience of participating in the tests has been relayed: the user said “to help me communicate with compensation”, and it went to check the order, check the flight dynamics, contact the passenger’s uniforms, and even sought $70 in cash for the vouchers that the platform would have given. This is not an isolated case: similar products are being introduced in the market, and the products of Muse in Meta, Cue in Manus, and a number of domestic manufacturers are being produced.
This series of signals is being validated, and Agent has moved from the demonstration phase to the day-to-day business and has begun to scale down. In May this year, Gartner projected that AZAI agent software spending would increase from $86.4 billion in 2025 to $206.5 billion in 2026 and $37.63 billion in 2027.
To scale, however, a more realistic premise needs to be met: during the last technological era, clouds were created to apply operating standards and their technological form was defined. Today, Agent is beginning to scale down and needs “another cloud”.
The difference is in the form of a task: an application to answer a question can close the request after returning the text; an Agent, who works for the user, waits for a response from the external system, generates documents in the interim and adjusts next step on the basis of feedback, users turn off the interface and the task is not necessarily over. When such missions operate in billion units, the cloud is no longer just a calculator and model, but an environment that allows Agent to continue to operate.
Internationally leading cloud manufacturers have begun to switch collectively. In its report “Agentic AI Cloud Titans in Asia & Oceania, 2026, the research institute Omdia rated AWS, Alibaba Cloud, Google Cloud and Microsoft Azure as four leaders of Agentic AIyun, who completed their judgments almost at the same time and began to adjust their technology depots. Google Cloud announced at the Next Congress that it would move from Claude Native to Agent Native, AWS introduced Bedrock AgentCore, Microsoft Azure's Founder Service was formally commercial, and Alibaba Cloud was more explicit about Agentic Cloud.
In the case of Alibaba Cloud, after the transition to Agenic Cloud was proposed in May this year, its technology vault began to adapt and integrate quickly, and in four months' time, at the Cloud Habitat Congress, AlibabaCEO U-wook said more directly, “Chips are responsible for the production of Token, clouds are responsible for sending Token to every place where it is needed”, and “AI clouds designed for models and for the continuous operation of Agent”.
In his judgement, the cloud would also have a network effect, and more users and Agent would join and further expand the network once the computing resources and tools were jointly optimized.
This provides an observational perspective similar to Metcalf ' s law, which describes the potential value of a connection using the relationship between network value and the number of users squared. The more tools Agent can use, the more business information he can get and the better he can work with one another, the more tasks he can undertake and the greater the economic value he can create.
At the same conference, Alibaba CloudCTO Li Fei further explained how the system was built, “one is production intelligence and one is creation of value”, which determined how modelling capabilities were produced and supplied, and the other answered how they were used in enterprises. On the other hand, the two were what Argentina Cloud needed to deliver.
Agent's size is on the ground, and the cloud is no longer just a resource.
Agent infrastructure could become an area of joint input by cloud manufacturers because it delivers not a platform interface that ends at once, but a run-through process.
The number of such tasks is growing. Gartner predicts that 33 per cent of enterprise software applications will be embedded in Agenic AI by 2028, and that this percentage will be less than 1 per cent by 2024, by which time at least 15 per cent of day-to-day work decisions will be AI agent-owned.
Markets give different paths. Some supermassive cloud manufacturers have made the operating environment and enterprise-level organization of AI agent a standard cloud product, while others have supplied these single points from sandboxes, running hours and Token. In the same direction, the cloud is being transformed into an infrastructure capable of carrying Agent.
For enterprises, the cost of using these off-the-shelf capabilities depends on the cost of building them. Self-building involves linking models and business systems, managing the state with permission, handling anomalies and establishing an evaluation mechanism for each change, and the more applied the maintenance of these links. The basic capacity provided by the cloud manufacturer to undertake this part of the work for the enterprise, which still needs to define its business rules and acceptance check results, does not have to re-establish the operating environment for each Agent.
Alibaba Cloud chose to be Agentic Cloud to bring these common works together into the cloud delivery.
This decision stems from insights into the needs of users, considering how models are produced and how enterprises construct and use AI agent. Li Fei Fei indicated that model capabilities determine the level of intelligent decision-making, Context provides the basis for decision-making and Harness provides assurance of implementation and continuity.
However, to achieve the Internet value of AI Yun mentioned at the beginning of the article, a single point is reliable enough. A model makes a miscalculation, which may be insufficiently reasoned or old data obtained; a plan is not implemented successfully for reasons that may arise in tool call or in the operating environment. Increasing model scores alone does not automatically solve other chain problems.
Alibaba Cloud thus extends capacity-building to model training and reasoning, the construction and operation of AI agent, and the organization and use of business data, which, compared to the procurement of a product, obtains an infrastructure that can work together.
How does Agenic Cloud catch it?
The business entrusted to Agent was a complete task, depending on how the model judged, context to provide accurate business information, and Harness to organize tool call and adjust steps based on implementation feedback. Anything that can't keep up could stop the mission. The solution to be addressed by Agenic Cloud is to align the three elements with the construction and operation and to support subsequent governance and iterativeity.
Li Fei has further dismantled the problem at the Cloud Habitat Congress and, in the context of client feedback and landing practices, he has consolidated the support needed at the enterprise level for Z42Z into four capacity pillars.
The first is the Global Cognitive Engine, which addresses the issue on which Agent based its judgement.
The quality of Agent ' s judgement depends to a great extent on what data it can see, not just how powerful the model is.
Businesses usually have a large amount of data that may not necessarily be directly available for use in missions, with information distributed in different systems, the same concept may be expressed differently and the state of business is changing. The model needs to know what is effective, what is relevant to the current mission and whether it has the right to use it.
Alibaba Cloud handles such needs through databases and large data capabilities, and OpenLake provides a global multi-mode data lake silo, unties engines and data, supports data and the same data as the AI engine controlled equal consumption. ApsaraLakebase is an integrated, mathematically integrated structure of a lake bank that combines data lakes with database capacity and supports data version management. The context project at Agent Context further collates the scattered information into the information required for the mission and can be used directly by Agent.
The data will also need to be updated on an ongoing basis, with a manual revision and a new business rule that could change the basis of judgement for the next mission, and the enterprise will therefore need to manage to include both raw information and experience gained during implementation.
The second is self-functioning and continuous evolution, which corresponds to Harness's primary responsibility.
Agent is not a single call that can be completed, with multiple tools and rounds of execution. Harness provided an implementation framework for tool call through an organizational model, arranged implementation steps and adjusted next step to the results.
Alibaba Cloud has built the platform AgentCore for the construction and operation of enterprise-level AI agent, standardizing and hosting the infrastructure that Agent needs to support long missions, failed retry, break-point recovery and walk-in implementation. The main capabilities include the segregation of sandboxes in the implementation environment, storage of storage status and documents, and monitoring to help teams understand the steps taken by the mission.
These capabilities address different issues, but need to work together: After the interruption of the mission, if documentation and state cannot be restored, Agent may be able to re-exercise the completed operation; if the call fails, it is difficult for the team to judge whether the model tips should be modified or the tools and environment adjusted if there is a lack of records.
The third is credible collaboration and governance, corresponding to the permission layer.
Agent uses the enterprise's systems and data, and the firm has to know who it is and what it can touch before it can really hand over the business to it.
When Agent enters the enterprise system, it needs to identify the identity and permission, on whose behalf the mission is carried out, which operations are allowed to be done autonomously, which must be identified, with specific rules, multiple Agents working together, and how the information is shared, and how responsibility is traced.
Alibaba Cloud's AI agent identity service Zgent Identity supports identity and authorization, combines security mechanisms to regulate the practical operation of Agent, while observable products assist businesses in analysing the process of calling and anomalies, and together determine the extent to which they can free Agent.
Li Fei Fei said, “Credibility, trust and capacity to pay are crucial, and I think the importance of this layer will be increasingly prominent”. It should be possible to check whether a mandate is correctly mandated, what tools are called upon and what results are produced.
The fourth is smart real-time decision-making, which addresses real access to business sites for the first three.
When business information reaches context in a timely manner, the model is critical and operational and governance mechanisms are in place to enable Agent to act in the actual state of business workflow.
The above-mentioned systematization requires a test of the bottom-level capability: model response is fast, reasoning costs are low, and Agent is called repeatedly in the business rhythm. Alibaba Cloud uses elastic support model training and reasoning, and the CPCS handles data reading and writing in large-scale training — training saves detection points, recovers after failure, and the speed of reading and writing determines the utilization of expensive computing resources, while the reasoning lines need to address issues such as the movement of resources and long context, which would not have been used if the request had been in the queue for too long.
The above-mentioned four capabilities are at different levels, and Li has emphasized that sandboxes and storage provide the conditions for implementation, Harness organizational tasks and feedback, and business data form context, which, if fully integrated into a single tool, cannot explain the different problems that need to be addressed in the actual works.
The project's on the ground.
The existence of the competency system depends on its performance in real business, which tests engineering.
Models give probabilistic judgements, and the same model is integrated into different systems, and the results can vary significantly; an enterprise needs a stable mission, from the organization of the data, the execution of the task, to the positioning and repair of the problem, each of which depends on engineering.
Three enterprises, three engineering challenges, are easier to explain than the chart why these capabilities need to be used in tandem.
The collaboration between MiniMax and Alibaba Cloud covers a complete link from model training to reasoning application.
Enhanced learning requires that the model itself enter the environment to change the code, compile, run the test - failed to change on the basis of feedback and succeeded in preserving the entire trajectory into the next round. The data thus moved from static to dynamic, and the demand for sandboxes was much more than traditional cloud computing: one million-class sandboxs were quickly loaded, the initial destruction was compressed to a few seconds and responded to a one-minute elastic fluctuations ranging from 100,000 to one million. Alibaba CloudAgent Sandbox can lift up, support and operate at a million-degree level per minute, activates about one to two seconds and heat loads within 100 milliseconds.
The assessment of whether the model is really strong is equally dependent on infrastructure. To assess the complexity of a model entering the real environment, each checkpoint pulls up a large number of sandboxes and repeats; the calculation and storage of instability does not distinguish between failure from a sandbox shaking or a model capability. Mini Max Research and Development Engineer and Developer Relations Manager Lee Yuan said that flexible cost control would allow us to know at a sufficiently low cost whether the model had become stronger or had a higher score.
Modelling capacity also needs to become a stable product. The essential difference between Agent and chat robots is whether or not to implement the environment and state — users write half of the code, intermediate files and context, all of which need to be fully preserved. MiniMax had previously introduced an object storage programme, which became unacceptable as the number of user files had increased and the waiting time had increased to a dozen seconds; after a few months of feedback demand to Alibaba Cloud, AgenticFSpublish, each user had an independent performance quota to mount Agent Sandbox directly. Li Yuan said that the most valuable was not to buy cloud products, but to have a cloud infrastructure that could quickly overlap with the team.
The Peng Group is the enterprise data base with Alibaba Cloud, which organizes decentralized business knowledge into context available to Agent.
As Agent enters the actual workflow process, the team needs to address the growing knowledge, support data space in different Agents and allow memory to cover more business information. This data base covers business knowledge and training needs, of which ApsaraLakebase is one of the pillars. Only when business rules and historical information are organized will it be possible for Agent to understand the request accurately and complete the operation.
In the design scene, the team provided Agent with historical asset and the evaluation process, as well as the designer ' s philosophy, with a single generation compliance rate of 83.3 per cent.
The problem of mission positioning and cost matching has been resolved throughout the life cycle with Alibaba Cloud through the development of an observable system and the optimization of Harness.
In a lifetime CAIO NUNE, the company-level office assistant Xiao Ying has served some 13,000 employees. When the scale of application is expanded, the team needs to know where the task is at and to avoid using an overcost model for simple requests. To this end, the Alibaba Cloud AI gateway and AgentLoop have built an observable and analytical system throughout their lives. On the observational basis of each reasoning, the Harness Project closed-ring “observation, assessment, optimization” was constructed, and models were selected according to the user's intent to match the calculation input to the task. At present, the mechanism has increased task success rate by 10 percentage points and the average model call cost by over 30 per cent.
AgentCore was used as a base for operations to bind the operation with sandbox isolation, safety testing and authorization. This increase in effect is a result of a systematic and joint adjustment of multiple links rather than a single upgrading of a particular product.
In addition, the operational capacity of the in-house applications has been made available throughout the life of Heil Yingkang as an Agent-accessible tool, with process experience deposited into skills, and the data generated from the application entered the enterprise data lake for re-use in other jobs. The reuse of this capacity is directly related to the network effects of AI clouds. When an operational capacity-building is completed, the value is not limited to the initial application if it can be called upon for more tasks. However, connectivity must be useful and information-sharing needed to be consistent with the permission border; a simple increase in the number of Agents would not guarantee a synchronized increase in business returns.
The three enterprises, each of which reduced the problem to the supply of work in the model operating environment, the construction of the enterprise context and the closure of the Harness project, needed to address different details, and in common, continued work in order for the model capacity to be used in a stable manner.
The economic value of Agenic Cloud is reflected here. The enterprise ' s aim is to make the work more reliable and economical. Alibaba Cloud has implemented the model, Harness and context in real business through product systems and engineering practices.
For enterprises, it was only when the delivery cycle had been shortened, task success rate had been raised, costs had been reduced and Token had acquired a clear value on the enterprise ' s books. (This article was first published on titanium medium APP)
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.