过去两年，AI Coding 几乎成为软件开发者的标配。无论是代码补全、函数生成，还是 Bug 修复、代码重构，AI 都在不断提高开发人员的工作效率。
但一个值得注意的现象是，开发者写代码的速度越来越快，企业的软件项目交付效率却未必同步提升。
美国软件工程数据分析与研发效能管理公司 Faros AI 分析了超过 1 万名开发者、1255 个软件工程团队的工作流数据，涵盖版本控制、代码审查、任务管理和 CI/CD 等系统。
尽管开发者完成了更多任务，研究却没有发现 AI 采用率与企业层面关键交付指标改善之间存在显著相关性。更有意思的是，PR 合并数量增加了 98%，代码审查时间却增加了 91%。 这反映出一个可能的瓶颈转移：代码生成加速之后，审查、测试和集成未必能跟上。
原因并不复杂。一个软件项目从提出需求到最终上线，除了编码，还涉及需求分析、产品设计、架构规划、测试验证、部署发布，以及后续维护。即使 AI 能让某个开发环节的效率提高 30%，也不意味着整个项目可以提前 30% 完成。
这也是当前 AI Coding 行业面临的一个新问题：当代码生成能力逐渐成为基础能力，下一步究竟应该优化什么？
10 月 26 日，网易智企正式发布 CodeWave SDD「AI 软件交付团队」，试图将 AI 的作用范围从单个开发者扩展到软件交付全流程。
与常规 AI Coding 工具不同，这款产品不再只为开发者提供代码生成能力，而是通过需求分析师、架构师、开发工程师、测试工程师等多个 Agent 分工协作，完成从需求理解到应用发布的完整流程。
网易智企副总经理、CodeWave 业务负责人王桐此前提出了一个判断：“我们现在跟企业聊，就不聊开发提效了，直接聊扩产能。”
按照网易智企披露的数据，该产品经过内部验证及超过 100 家企业的实际项目验证，能够实现约 3～5 倍的产能提升。
但相比提效数字，更值得关注的是其背后反映出的行业变化：AI Coding 的竞争重点，正在从帮助程序员更快写代码，向如何组织 AI 完成完整的软件工程任务延伸。
AI Coding 的瓶颈，可能不在代码生成
王桐谈到，网易内部已经广泛使用各类 AI Coding 工具，开发单点环节获得了约 20%—～0% 的效率提升。
但如果从企业整体经营角度衡量，这种改善可能被稀释至个位数百分比。
对于软件开发服务商而言，这种差异尤为明显。
软件项目的交付成本并不只由程序员写代码的时间决定。需求反复变更、不同岗位之间的信息传递、测试返工、历史系统维护，以及最终客户验收，都可能消耗大量资源。
特别是面对大型企业项目，一份需求文档可能包含上百页内容，既有业务流程，也有界面设计、权限规则和复杂的数据关联。即使模型能够快速生成代码，也很难仅靠一次提示就准确理解全部要求。
开发者通常需要先手动拆解需求，再将不同任务分批交给 AI Coding 工具，最后自行整合代码、检查一致性并解决冲突。
这意味着，AI 虽然承担了部分编码工作，但需求拆解、任务组织和交付管理仍然需要大量人工参与。
CodeWave SDD 试图改变的正是这一工作方式。
“我们的下一代，不是让人用 AI Coding 工具，而是直接让 Agent 员工去用 AI Coding 工具。”王桐表示。
具体而言，开发者不再需要围绕单个工具持续编写提示词、拆分任务和检查生成结果，而是将完整需求交给一个由多个专业 Agent 组成的软件交付系统，由其自主规划、协作和执行。
当然，这并不意味着 AI Coding 工具本身已经失去价值。相反，代码生成工具仍然是这类系统的重要基础，只是人机交互和任务组织的层级发生了变化。
一支 AI 软件交付团队，如何真正完成复杂项目？
从技术实现看，CodeWave SDD 的主要特点是试图通过统一的工程规范和执行环境，将不同 Agent 组织成一条完整的软件交付流水线。其中有两个值得关注的技术设计：
首先是规格驱动开发（Spec-Driven Development，SDD）。在传统开发流程中，需求文档往往存在不完整甚至相互矛盾的情况。人类工程师可以通过会议、沟通和经验补充缺失信息，但 AI 如果直接基于模糊需求生成代码，容易出现功能遗漏、理解偏差等问题。
为减少这种不确定性，CodeWave SDD 首先由需求分析师 Agent 读取需求文档，并识别其中的文本、表格、流程图、原型图等多模态信息。
对于不明确的需求，Agent 会主动向人提问，完成澄清后，再使用 EARS 等结构化需求描述方法生成相对明确的规格文档（Spec）。
随后，设计师 Agent 生成原型，架构师 Agent 根据规格设计技术方案并拆解任务，前后端开发 Agent 负责代码实现，测试 Agent 则依据规格生成测试用例并执行验证。
王桐强调：“严肃的任务，我们是不希望模型做发挥的。”
这句话实际上点出了企业软件开发与一般 AI 内容生成的重要区别。对于创意写作，模型的自由发挥可能带来惊喜；但对于金融、政企等业务系统，权限控制、数据结构和操作逻辑必须符合明确要求。任何未经确认的功能增减，都可能引入业务风险。
因此，CodeWave SDD 采用了将大型需求拆解为大量细颗粒度任务、分批执行并持续检查规格一致性的方式。
另一个关键设计，是企业级 Harness 工程环境。
如果说多个 Agent 相当于软件团队中的不同成员，那么 Harness 则负责为它们提供统一的运行环境、执行约束与协作机制。
王桐透露，网易为这套产品投入了接近 100 人的研发团队，打磨了大半年。其中约 70% 的研发资源和时间用于构建 Harness 环境，其余约 30% 投入多 Agent 协作相关工作。与运行在个人终端上的 AI 编程工具不同，CodeWave SDD 将 Agent 团队部署在企业级集群环境中。
这套环境需要持续检查 Agent 是否正常工作、任务是否超时、不同 Agent 之间的通信是否中断，以及每个阶段的交付物是否符合既定规格。
如果检查未通过，系统会阻止任务进入下一阶段，并执行相应的修复流程。这一设计对应着 AI Agent 落地企业生产环境的现实难题：模型能否完成某项任务，与一个系统能否连续运行数十小时、处理异常并稳定交付，并不是一回事。
对于企业而言，后者往往直接关系到产品是否真正可用。
从百页 PRD 到老系统维护，AI 能承接多少实际工作？
除了自动生成新应用，CodeWave SDD 还针对存量系统改造设计了代码理解环节。王桐介绍，面对已有系统时，平台会先读取历史代码，通过 AST 语法树和 LSP 语义服务建立代码地图，梳理程序结构与代码之间的关联，再根据新增需求开展迭代。
这项能力针对的是企业软件开发中的一类典型痛点：大量开发任务并非从零构建新系统，而是在业务逻辑复杂、文档可能不完整的旧系统上持续修改。
如果 Agent 无法理解已有代码和依赖关系，即使能够生成新功能，也可能破坏原有系统。
为了展示上述效果，网易智企选取某省级担保公司的核心业务系统作为案例，该项目需求文档超过 100 页，涉及业务台账、数据分析、复杂检索等多个模块。
按照网易智企展示的对比结果，在使用相同需求与模型的条件下，CodeWave SDD 基本还原了需求中的功能模块，而参与对比的一款 IDE 类工具遗漏了部分子菜单、检索能力及页面交互功能。
在这项任务中，需求分析和架构设计 Agent 还生成了 200 多份规范文档，形成从需求到技术方案的过程记录。这些结果说明，在复杂需求的拆解与组织方面，完整的工程化流程可能比单纯增加模型上下文更有价值。
不过，单个项目的展示尚不足以证明不同工具在所有软件开发任务中的能力差异。项目复杂程度、提示方式、任务拆解机制和验证标准，都可能影响最终结果。
对于 AI 生成代码，另一个需要关注的问题是后续维护。
王桐表示，CodeWave SDD 通过细颗粒度 Spec、过程校验和分步执行约束 Agent 行为，并保留需求文档、架构设计、完整源代码、测试用例及报告等交付资产。
按其披露的内部测试数据，P0 级重要测试用例的一次通过率约为 90%，P1、P2 级测试用例约为 80%。
此外，针对企业代码和业务数据安全问题，网易智企表示，平台支持本地部署，可接入企业自身的知识库、工具和模型，并在开发、测试流程中集成安全检测机制。
当 AI 承担开发，软件工程师的价值会转向哪里？
技术变化最终也会影响组织结构。
根据网易智企披露，目前 CodeWave SDD 已在超过 100 家企业开展项目实践，其中约 80% 为软件开发服务商，涉及管理系统、移动应用、数据分析等场景。
过去，开发团队往往将大量人力配置在编码和具体技术实现环节。随着 Agent 能够承担越来越多执行性工作，需求理解、技术方案评审、质量验收等环节的重要性可能进一步上升。
王桐提出了一种组织调整思路：原来由一名产品经理配合十名开发人员完成的工作，未来可能转变为五名产品人员与五名开发人员协同，由 Agent 承担更多具体执行任务。
这当然不是适用于所有项目的固定比例，但它反映了 AI Coding 对软件开发组织提出的新要求。
此外，王桐还谈到，企业需要让开发人员逐渐掌握业务需求梳理和交付验收能力。他认为，FDE（前沿部署工程师）的核心不只是掌握 Agent 开发工具，更在于能够理解客户业务，把模糊诉求转化为明确需求，并对最终交付结果负责。
事实上，AI 参与开发越深入，人类对需求和结果的判断就越重要。
当大量代码不再由人逐行编写，企业必须回答新的问题：谁来决定系统应该做什么？谁来判断 AI 生成的功能是否符合业务要求？系统出现问题之后，又由谁承担维护和交付责任？
至少在当前阶段，这些责任并不能随着代码生成自动化而一并转交给 AI。
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。
In the last two years, AI Coding has almost become the standard for software developers. Whether code completion, function generation, or Bug repair, code re-engineering, AI is constantly improving the efficiency of developers.
It is worth noting, however, that developers are writing codes more and more quickly, while the efficiency of enterprise software project delivery is not necessarily increased simultaneously.
The United States software engineering data analysis and development effectiveness management company Faros AI analysed workflow data for over 10,000 developers and 1255 software engineering teams, covering systems such as version control, code review, task management and CI/CD.
While developers have done more, the study has not found a significant correlation between AI adoption rates and improvements in key delivery indicators at the enterprise level. And more interestingly, the number of PR mergers has increased by 98 percent, while code review times have increased by 91 percent. This reflects a possible shift of bottlenecks: review, testing and integration do not necessarily keep pace after code generation has accelerated.
The reasons are not complex. A software project, from demand generation to end-line, involves, in addition to coding, needs analysis, product design, architecture planning, testing validation, deployment of publish and follow-up maintenance. Even if AI can increase the efficiency of a development chain by 30%, it doesn't mean that the whole project can be completed 30% ahead of schedule.
This is also a new problem facing the AI Coding industry: What should next step do when code generation becomes basic?
On October 26th, the InternetIntelligence Corporation officially publish CodeWave SSD "AI Software Delivery Team" attempted to extend the role of AI from a single developer to a full software delivery process.
Unlike the regular AI Coding tool, the product no longer only provides code generation capability for developers, but rather completes the process from demand understanding to the application of publish through multiple Agent division of labour collaborations such as demand analysts, architects, development engineers and test engineers.
The Deputy Managing Director of the Internet-based Intelligence Enterprise and Officer-in-Charge of the CodeWave business, Wang Kiri, made a judgment: "We talk to business now, we don't talk about development, we talk about capacity enhancement."
According to data disclosed by Internet-based ingenuity, the product has been validated internally and by actual projects in more than 100 enterprises, which has resulted in about three to five times the increase in capacity.
But it is more interesting to note the changes in industry that are reflected behind this figure: the competition focus of AI Coding is extending from helping programmers write codes faster to organizing AI to complete software engineering tasks.
Al Coding bottlenecks, probably not code generation
Wang Kui said that a wide range of AI Coding tools have been used within the Internet, and the development of single-point links has resulted in about 20% - 0% efficiency improvement.
However, this improvement may be diluted to a single-digit percentage, measured in terms of the business as a whole.
This difference is particularly evident for software development service providers.
The cost of the delivery of software items is not determined only by the time the programmer writes the code. Repeated changes in demand, transmission of information between different jobs, back-to-work testing, historical system maintenance and the final client, acceptance check, can all consume considerable resources.
In particular, in the case of large enterprise projects, a demand file may contain hundreds of pages of both business processes and interface design, permission rules and complex data linkages. Even if the model is able to produce the code quickly, it is difficult to understand all the requirements with a single reminder.
Developers usually have to dismantle the needs manually before handing over different tasks to the AI Coding tool, and finally consolidate their own codes, check consistency and resolve conflicts.
This means that although AI has undertaken part of the coding work, demand dismantling, task organization and delivery management still require significant manual involvement.
CodeWave SDD is trying to change the way it works.
"Our next generation, instead of letting people use the AI Coding tool, let Agent's employees use the AI Coding tool." The Wang Yi indicates.
Specifically, the developers no longer need to produce the results of prompt, split missions and inspections on an ongoing basis around individual tools, but instead hand over the complete requirements to a multi-professional Agent software delivery system for autonomous planning, collaboration and implementation.
Of course, this doesn't mean that the AI Coding tool itself has lost value. On the contrary, the code generation tool remains an important foundation for such systems, except for human interaction and changes in the level of mission organization.
How can an AI software delivery team really complete complex projects?
In terms of technology, the main feature of CodeWave SDD is the attempt to organize different Agent into a complete software delivery line through a unified engineering code and implementation environment. There are two technical designs that deserve attention:
The first is specification-driven development (Spec-Driven Development, SDD). In traditional development processes, demand files are often incomplete or even contradictory. Human engineers can supplement missing information through meetings, communication and experience, but AI is vulnerable to functional omissions, misperceptions, etc. if it generates codes based directly on vague demand.
In order to reduce this uncertainty, CodeWave SDD starts with Agent, a needs analyst, reading the demand document and identifying the text, tables, flow charts, prototypes, etc.
In the case of undefined needs, Agent takes the initiative to ask people questions, and when clarifications are completed, the relatively clear specification document (Spec) is generated using structured demand description methods such as EARS.
Subsequently, the designer Agent generated the prototype, the architect Agent designed the technical program and dismantled the task according to the specifications, the front-end development of Agent was responsible for the code realization, and the test Agent was based on the specifications to generate test examples and perform validation.
Wang Yi stressed, "We don't want models to work on serious tasks."
This sentence actually points out the important difference between enterprise software development and general AI content generation. For creative writing, the free exercise of models may be a surprise; but for business systems such as finance, government and industry, permission controls, data structures and operating logic must meet clear requirements. Any unrecognized functional gain or loss may introduce operational risk.
Thus, CodeWave SSD uses the method of disassemblying large demand into a large number of fine particle size missions, batch execution and continuous inspection of specifications consistency.
Another key design is the enterprise-level Harness engineering environment.
If multiple Agents are equivalent to different members of the software team, Harness is responsible for providing them with a unified operating environment, enforcement constraints and collaborative mechanisms.
Wang Kiri has revealed that the Internet has invested nearly 100 people in the development of this product, which has been grinding for six months. About 70% of the R & D resources and time are devoted to the construction of the Harness environment, while the remaining 30% are heavily involved in Agent collaboration. Unlike the AI programming tool running on personal terminals, CodeWave SSDD deploys the Agent team to the enterprise-level cluster environment.
This set of environments requires an ongoing inspection of whether Agent is working properly, whether the mission is timeout, whether communication between different Agents has been interrupted, and whether each stage of delivery meets the established specifications.
If the inspection fails, the system prevents the task from entering the next phase and implements the corresponding restoration process. This design corresponds to the real challenge of the production environment of AI Agent's landed enterprise: it is not the same thing as whether the model is capable of performing a particular task, running a system for dozens of consecutive hours, handling anomalies and stabilizing delivery.
For enterprises, the latter tend to be directly related to the actual availability of products.
From 100 pages of PRD to old system maintenance, how much actual work can AI take on?
In addition to automatically generating new applications, CodeWave SDD also designed code understanding links for stock system adaptation. Wang Kiri explained that, in the face of existing systems, the platform would first read historical codes, create code maps through the AST syntax tree and LSP syntax service, combe the links between the program structure and the code, and then perform an iterative process based on additional needs.
This capability is aimed at a typical set of pain points in enterprise software development: a large number of development tasks do not build new systems from zero, but are continuously modified on old systems that have complex business logic and potentially incomplete documentation.
If Agent does not understand existing codes and dependencies, even if it is able to generate new functions, it can damage the old system.
In order to demonstrate the above effects, the web-based ingenuity has selected the core business systems of a provincial guarantee company as a case in which the project requires more than 100 pages of documentation, covering several modules such as business desk accounts, data analysis and complex retrieval.
According to the comparison presented by the web-based ingenuity, under the same requirements and models, the CodeWave SSDD essentially restored the functional module in the demand, while the IDE-type tool involved in the comparison omitted some submenu, searchability and page interaction.
In this task, Agent also generated more than 200 normative documents to document the process from demand to technical options. These results suggest that complete engineering processes may be more valuable than simply adding model context in terms of decomposition and organization of complex needs.
However, the demonstration of individual projects is not sufficient to demonstrate that different tools differ in capacity across all software development tasks. The complexity of the project, the mode of alert, the task dismantling mechanism and the certification criteria may affect the final outcome.
Another concern for AI generation codes is follow-up maintenance.
Wang Yi indicated that the CodeWave SDD delivers assets through fine particle size Spec, process validation and step-by-step execution of Agent behaviour, and maintains demand documentation, architecture design, complete source code, test examples and reports.
According to the internal test data disclosed by it, the pass rate for major tests at the P0 level is about 90 per cent and the pass rate for tests at the P1 and P2 levels is about 80 per cent.
In addition, in response to enterprise code and business data security issues, INEI stated that the platform supported local deployment, had access to the enterprise ' s own knowledge base, tools and models, and had integrated security detection mechanisms in the development and testing process.
When AI undertakes to develop, where will the value of software engineers turn?
Technological changes ultimately affect organizational structures.
According to Internet-based ingenuity disclosure, the CodeWave SSD is currently operating project practices in over 100 enterprises, of which approximately 80 per cent are software development service providers, involving scenarios such as management systems, mobile applications, data analysis, etc.
In the past, development teams have often deployed large numbers of people into the coding and technical realization chain. As Agent is able to take on an increasing number of implementation tasks, the importance of components such as demand understanding, technical programme evaluation, quality acceptance check may further increase.
Wang Kiri put forward an organizational adjustment idea: the work that had been done by a product manager in conjunction with the 10 developers could be transformed into five producers working with five developers, with Agent taking on more specific tasks.
This, of course, is not a fixed proportion applicable to all projects, but it reflects the new requirements of AI Coding for software development organizations.
In addition, Wang Yi spoke of the need for enterprises to develop the ability of developers to prepare their business needs and deliver acceptance check. In his view, the core of the FDE (Front Deployment Engineer) was not just the Agent development tool, but the ability to understand client operations, to translate vague claims into clear needs and to be held accountable for the final delivery results.
Indeed, the more AI is involved in development, the more important it is for humans to judge needs and outcomes.
When a large number of codes are no longer line-by-line, enterprises have to answer new questions: who decides what the system should do? Who's gonna judge if the AI generation function meets business requirements? Who is responsible for maintenance and delivery after problems with the system have arisen?
At least at this stage, these responsibilities cannot be transferred to AI as code generation is automated.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.