智东西（公众号：zhidxcom）
作者 | 毕伟豪
编辑｜心缘
智东西10月26日报道，刚刚，网易智企发布了面向企业的新一代AI Coding产品CodeWave SDD（AI软件交付团队），核心在于通过全自主Agent软件开发，实现企业整体产能提升3到5倍。
网易智企将CodeWave SDD定义为AI Coding的下一代产品形态：从人使用工具，转向Agent调用工具完成软件开发。
按照网易内部《AI落地成熟度标准》，这属于L4级别的自动驾驶，即AI端到端完成任务，人类主要负责需求梳理和交付验收。
为此，网易智企搭建了覆盖软件开发全流程的企业级Harness工程环境，让多个分工不同的Agent协同工作。
据网易智企介绍，CodeWave SDD目前已通过超过100家企业的实操验证。
一、让Agent实现“自动驾驶“，企业产能提高300%
关于新形态AI软件交付团队，网易智企副总经理、CodeWave业务负责人王桐给出了一个很形象的解释：不是让人用AI Coding工具，是让Agent去使用AI Coding工具。
王桐认为， Claude Code等AI编程工具，本质上是辅助人类工作的副驾驶，CodeWave SDD是在原有的AI Codeing工具之上叠加一层，让多个Agent员工在统一平台去操控工具，实现“自动驾驶“。
CodeWave SDD的核心结构分为两层，底层是一个跑在企业集群上的Harness工程环境，覆盖需求、原型、架构方案、开发、测试、部署，整条软件开发链路。
上层则是按软件开发环节分设的不同角色Agent，包括需求分析、设计、架构师、前端与后端开发等。核心概念是让不同Agent去调用AI编程工具执行相应环节的任务。
这套架构的优势最终反映到的是产能。按照网易智企的说法，之前企业使用AI编程工具所能实现的是百分比的效率提高，但使用CodeWave SDD，企业可以让Agent承担软件开发任务，直接扩大产能。
二、多Agent协同开发，人类只负责把关
CodeWave SDD的运转依赖一套SDD框架，也即规格驱动的软件开发。
这条开发流水线共分为6个部分。第一个部分是需求分析师Agent，它可以接收长上下文的需求，并接收多模态输入，对于需求模糊的部分，还会通过提问来确认需求。
第二部分则是通过EARS语法将需求改成没有歧义的规格文档，提供给下游架构师Agent。
第三是设计师Agent，这个Agent主要提供产品原型。在Codewave SDD这个平台上，业务方也可以进行需求追加，或者在原型图上进行圈选修改。
第四段由架构师Agent写技术方案，方案里包含开发任务怎么拆分并交给开发Agent。
第五段是前端和后端开发Agent进行具体软件开发的实现，在软件开发过程中可以实时预览开发效果。
整个开发的最后一环是测试Agent，它会先生成完整用例，再执行测试，然后出具报告，最后推送Git仓库，支持分支管理和多人协作，并可以上传到平台提供的环境中。
对于存量系统，在这些流程之前，有一个叫代码地图的环节。企业可以导入源码或Git链接，Agent会解析前后端和配置文件，将原有的逻辑关系整理成图表，然后再根据需求进行开发，这点主要是应用在旧系统的升级和迭代。
整个开发环节，人类只需要进行需求的梳理以及交付验收即可，全流程都可以通过AI端到端完成。
但这两个环节事实上也非常重要。行业专家在这两个环节的价值是将要求翻译成需求，并仔细核验AI生成的产物。王桐在接受采访时提到，AI是不能帮人去背锅的，也即交付物不能直接甩给业务方，需要人去核验。
在组织层面，网易内部已经因此出现了变动，BA和产品岗的占比正在提升，产品开发成员部分转去更懂业务和验收的位置。
在安全方面，王桐介绍，CodeWave SDD主要从三个层面进行防护：
一是支持整套Agent团队和企业Harness平台本地部署，方便与企业代码库、知识库及工具库集成；
二是在编码和测试环节内置安全漏洞检测体系，产出通过检测后才能交给下一个Agent，企业也可以补充4A认证等个性化安全要求；
三是采用步进式调用，将复杂需求拆分成多个任务分批处理，避免一次性上传完整核心资产，同时根据不同Agent的职责组合使用不同模型，以降低潜在风险。
三、与IDE相比优势明显，P0用例一次通过率90%
现场演示中，王彤展示了CodeWave SDD与主流IED类工具的实际开发对比效果。
测评使用的是某省级担保公司的核心业务系统开发，PRD超过了100页。同样的需求和模型，开发过程中，CodeWave SDD基本还原了功能模块，而IDE则是仅做了一级菜单，子菜单全部丢失。
从页面细节上看，台账的概览页面，CodeWave SDD的标签页划分指标卡和BI图表都是根据需求构建的。而IDE则出现了各种问题，比如没有按视角分标签页，柱状图和双色折线图缺失等等。
对此，王桐提出，造成这种情况的原因是上下文需求超过了IDE的承接范围，一些复杂的任务必须由专人拆解，再提供给IDE类工具，因此它并不是自动开发。
而CodeWave SDD可以接受模糊需求，然后根据后续的提问进行需求的设计与完善，实现所谓的自动驾驶。
网易披露，CodeWave SDD的P0级别测试用例一次通过率为90%，P1、P2约为80%。，不过这一指标反映的是测试用例首次执行的通过情况，并不等同于整个项目的交付成功率。
四、不止产品，还包括服务，三步避免“屎山代码”
此前，王桐在接受智东西采访时提到，CodeWave SDD除了产品本身，还提供模型推荐、本地算力支持、客户成功和人才培训等服务。
在模型选择上，每个Agent都可以配置独立模型。网易会在新模型发布后进行全场景测评，并在24小时内形成报告，为不同Agent推荐合适的模型。
对于有本地算力需求的国央企、银行等客户，产品也支持本地部署，并提供模型选择和配置建议。
服务方面，网易会为企业提供组织转型和人才培养服务。网易会培训开发人员梳理需求、把关验收，帮助他们从单纯的技术执行转向理解业务、与客户沟通的角色。
王桐认为，FDE（前沿部署工程师）不只是会搭建Agent，更重要的是能在客户现场收集零散的业务要求，结合业务理解梳理出真正的需求，再交给AI执行，并对产出进行验收和修改。
面对智东西提出AI生成代码可能难以维护的“屎山代码“问题，王桐用三句话解释了CodeWave SDD在这方面的解决方案。
第一，以细颗粒度的Spec作为统一标准。 需求规格需要足够明确，所有Agent都以此为依据开展工作，确保需求、开发和测试有共同的判断标准。
第二，通过Harness环境进行硬性校验。 从原型到代码，各阶段产出都要检查是否符合Spec。未通过检查，就不能进入下一阶段。
第三，将大任务拆成小任务，分批执行。 一个复杂需求可以拆成数百个细分任务，每次只执行其中一部分，完成后先对照Spec检查，再继续下一批，减少任务过大导致的偏差。
王桐称，网易希望通过这三种方式，让Agent生成的代码更符合规范，也更便于后续维护和迭代。
结语：网易新招，能否啃下企业AI市场？
网易此次发布，瞄准的是企业AI软件交付的市场，但与常规AI Coding工具让开发者更快产出代码不同的是，网易所做的工作，是让Agent承担更多原本由人完成的工作。
CodeWave SDD想做的自动化，是让Agent占据部分员工的生态位置，在员工数量做到提升，从而影响产能。这种形式能否跑通，不仅取决于基模能力，与需求处理、交付质量和维护能力紧密相关。
按照网易的设想，企业可以通过增加Agent承担更多开发工作，让人把精力转向需求梳理、架构评审和交付验收，进而改变团队分工和产能结构。
不过，软件开发涉及复杂的业务逻辑和长期维护，从实际情况上看，想实现Agent“自动驾驶“还不现实，人类仍需要在需求和验收上严格把关。
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。
Intellectual (WeChat Official Account:zhidxcom)
The author.
Edit Thread
On October 26th, J.I.A. reported that, just now, I.I.E. publish had a new generation of company-oriented I.I.C. CodeWave SDD (A software delivery team), with the core of which is the three-to-five-fold improvement of corporate capacity through the development of fully-owned Agent software.
InternetIntelligence defines CodeWave SSD as the next generation of products in AI Coding: from human use of tools to Agent call tools to complete software development.
In line with the Internet-based internal " ASI maturity standard ", this is a L4-level automatic driving, i.e., an AI-to-end mission, where humans are primarily responsible for combing needs and delivering acceptance check.
To this end, the Internet-based Ingenuity Initiative has created an enterprise-level Harness engineering environment covering the entire process of software development, which brings together multiple and divided Agents.
According to Internet Intel, CodeWave SDD has now been certified through the exercise of more than 100 businesses.
I. “Auto-driving”, 300 per cent improvement in enterprise capacity
With regard to the new form of AI delivery team, Wang Yi, Deputy Director-General of NetIge and Head of the CodeWave Operations, gave an image: it was not an AI Coding tool, but an AI Coding tool.
Wang Qi believes that the AI programming tools such as Claude Code are essentially co-pilots that support human work, and that CodeWave SDD is superimposed on the original AI Codeing tool, allowing multiple Agent employees to manipulate the tool on a unified platform and achieve autopilot.
The core structure of CodeWave SSD is divided into two layers, with a Harness engineering environment running on enterprise clusters, covering needs, prototypes, architecture programmes, development, testing, deployment, and the entire software development chain.
The upper tier is the different roles of Agent by the software development chain, including needs analysis, design, architecture, front-end and back-end development. The core concept is to get different Agents to use the AI programming tool to carry out the corresponding tasks.
The strength of this architecture ultimately reflects capacity. According to Internet-based ingenuity, the former use of the AI programming tool had resulted in an increase in percentage efficiency, but the use of the CodeWave SDD allowed Agent to undertake software development tasks and directly expand its capacity.
ii. MultiAgent development. Humans are only responsible for closing.
CodeWave SDD operates on a set of SDD frameworks, i.e. specification-driven software development.
The development line is divided into six parts. The first component, Agent, is a demand analyst who can receive demand for long context and receive multi-modular input, and for vague demand, needs are identified by asking questions.
The second part is the conversion of demand into a non-differentiated specification document through the EARS syntax, which is made available to Agent, a downstream architect.
And the third is the designer Agent, who mainly provides the prototype. On the Codewave SSD platform, the operator may also make additional requests, or re-choice changes on the prototype.
The fourth section is the Architect Agent Writing Program, which includes the development task of how to split and gives it to Agent.
The fifth paragraph is the achievement of specific software development by developing Agent at the front and back ends, allowing real-time previews of development effects during the software development process.
The last link to the entire development is the testing of Agent, which will be used in its entirety by the gentleman, who will then perform the test, then produce the report, and eventually deliver it to Git's warehouse, support branch management and multi-person collaboration, and can be uploaded into the environment provided by the platform.
For stock systems, these processes are preceded by a link called code map. Enterprises can import source codes or Git links, and Agent will analyze the back, back and back ends and configuration files, organize the original logical relationships into charts and then develop them according to demand, mainly for upgrading and iterative application in the old system.
The entire development chain, where humans simply need to comb and deliver acceptance check, can be completed through the AI-to-end process.
But these two links are also, in fact, very important. The value of industry experts in these two links is to translate the requirements into demand and to scrutinize the products generated by AI. In an interview, Wang Kiri stated that AI could not help people to take the fall, i.e. delivery could not be dumped directly to the operator and needed to be verified.
At the organizational level, changes have taken place within the ease of the network, with the share of BA and product placements increasing, and product development members shifting in part to a better understanding of business and acceptance check positions.
In terms of security, Wang Kiri, the CodeWave SSDD works on three levels:
One is to support the deployment of the entire Agent team and enterprise Harnesssping Z89Z to facilitate integration with the enterprise code bank, knowledge bank and tool bank;
Second, a security gap detection system is built into the coding and testing chain, with the output being tested before being handed over to the next Agent, and enterprises can add personalized security requirements such as 4A certification;
Third, a step-by-step approach is used to process complex needs split in multiple task groups, avoiding one-time upload of full core assets, while using different models based on different combinations of Agent ' s responsibilities to reduce potential risks.
III. A clear advantage in relation to IDE, with a 90 per cent pass rate for P0 cases
In a live demonstration, Wang Xin demonstrated the practical contrast between the CodeWave SDD and the development of mainstream IED-type tools.
The evaluation used the core business system development of a provincial guarantee company, and the PRD exceeded 100 pages. The same needs and models were developed with the functional modules largely restored by the CodeWave SSD, while the IDE was only able to make the first-order menu, and the sub-menu was all lost.
From the page details, the overview page of the desk account, the CodeWave SSD tab page split the indicator card and the BI chart are based on demand. IDE, on the other hand, had problems, such as the absence of perspective tabs, the absence of column charts and double-colour circuits.
In response, Wang Yi argued that this was due to the fact that context requirements were beyond the reach of IDE, and that complex tasks had to be detached and then provided to IDE-type tools, so that it was not automatically developed.
CodeWave SSD, on the other hand, can accept vague demand and then design and refine it in the light of subsequent questions to achieve so-called automatic driving.
The easy disclosure of the CodeWave SDD test at the P-0 level is 90 per cent pass at a time and about 80 per cent at the P1 and P2 levels. This indicator, however, reflects the adoption of the test example for the first time and is not equivalent to the delivery of success rate for the project as a whole.
IV. Not only products, but also services, three steps to avoid the "Shit Mountain Code"
Previously, Wang Yi mentioned in an interview with intellectuals that, in addition to the product itself, the CodeWave SDD offered modeling, local computing support, customer success and talent training.
In model selection, each Agent can configure an independent model. The web easer conducts the whole scene assessment after the new model publish and produces reports within 24 hours to recommend suitable models for different Agents.
For clients such as national firms, banks and others with local computing needs, products also support local deployment and provide model selection and configuration advice.
In terms of services, the network provides organizational transformation and talent development services for enterprises. The Network trains developers to combe needs and turn off acceptance check to help them move from purely technical implementation to understanding business and communicating with clients.
Wang Kiri believes that the FDE (Front Deployment Engineer) is not only capable of building Agent, but also, and more importantly, of collecting scattered business requirements at the client ' s site, combing out real needs in the context of business understanding and handing over to AI for implementation and acceptance check and modifications of outputs.
In response to a question from intellectuals about the "Shit Mountain Code" that may be difficult to maintain, Wang Qi explained the CodeWave SDD solution in this regard.
First, Spec of fine particles is used as a uniform standard. The specification of requirements needs to be clear enough, and all Agents work on this basis to ensure that there are common standards of judgement for needs, development and testing.
Second, a hard check through the Harness environment. From prototype to code, all stages of output are checked for spec. Without the inspection, the next stage cannot be reached.
Thirdly, large tasks should be broken down into smaller tasks and carried out in batches. A complex requirement can be broken down into hundreds of sub-tasks, only part of which is performed each time, checked against Spec before continuing the next batch, reducing the bias caused by excessive tasks.
According to Wang Kiri, it is easy to hope that the codes generated by Agent will be more in line with the norms, and more easily followed up and iterative.
Concluding remarks: Is it possible to eat the corporate AI market?
In contrast to the conventional AI Coding tool, which allows developers to produce codes faster, the work done by the Internet is to get Agent to take on more human work.
The automation that CodeWave SSD wants to do is to allow Agent to occupy the ecological position of part of the workforce and to increase the number of employees, thereby affecting productivity. The viability of this form depends not only on the capability of the model, but also on the ability to process, deliver quality and maintain demand.
Under the easy-to-grid scenario, firms could take on more development work by adding Agent, which would shift their energies to demand-readiness, architecture review and delivery of acceptance check, thereby changing the division of labour and the structure of production capacity.
However, software development involves complex business logic and long-term maintenance, and in practical terms it is not realistic to achieve Agent's "automated driving" and humans still need to be strictly closed on demand and acceptance check.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.