NetEase Zhigi Releases CodeWave SDD, Using Multi-Agent Collaboration to Complete Software Development

On October 26, Zhi Dong Xi reported that NetEase Zhigi has just released CodeWave SDD (AI Software Delivery Team), a new-generation AI Coding product for enterprises, whose core is to achieve a 3 to 5 times increase in overall enterprise capacity through fully autonomous Agent software development.

NetEase Zhigi defines CodeWave SDD as the next-generation product form of AI Coding: shifting from humans using tools to Agents calling tools to complete software development.

To this end, NetEase Zhigi has built an enterprise-level Harness engineering environment covering the entire software development process, enabling multiple Agents with different divisions of labor to work collaboratively.

According to NetEase Zhigi, CodeWave SDD has already been validated through practical operations at more than 100 enterprises.

1. Enabling Agents to achieve "autonomous driving," boosting enterprise capacity by 300%

Regarding the new form of AI software delivery team, Wang Tong, deputy general manager of NetEase Zhigi and head of the CodeWave business, gave a vivid explanation: it is not about letting people use AI Coding tools, but about letting Agents use AI Coding tools.

Wang Tong believes that AI programming tools such as Claude Code are essentially copilots assisting human work, while CodeWave SDD adds a layer on top of the original AI Coding tools, allowing multiple Agent employees to operate the tools on a unified platform and achieve "autonomous driving."

The core structure of CodeWave SDD is divided into two layers. The bottom layer is a Harness engineering environment running on an enterprise cluster, covering requirements, prototyping, architecture solutions, development, testing, and deployment—the entire software development chain.

The upper layer consists of different role-based Agents set up according to software development stages, including requirements analysis, design, architect, and front-end and back-end development. The core concept is to have different Agents call AI programming tools to execute the tasks of the corresponding stages.

The advantage of this architecture is ultimately reflected in capacity. According to NetEase Zhigi, what enterprises could achieve previously with AI programming tools was a percentage-based efficiency improvement, but with CodeWave SDD, enterprises can let Agents undertake software development tasks and directly expand capacity.

2. Multi-Agent collaborative development, with humans only responsible for oversight

The operation of CodeWave SDD relies on a set of SDD frameworks, namely specification-driven software development.

This development pipeline is divided into 6 parts. The first part is the requirements analyst Agent, which can receive long-context requirements and multimodal input, and for ambiguous parts of the requirements, will also confirm them by asking questions.

The second part is to convert the requirements into unambiguous specification documents through EARS syntax and provide them to the downstream architect Agent.

The third is the designer Agent, which mainly provides product prototypes. On the CodeWave SDD platform, business parties can also add requirements or make circled modifications on the prototype diagram.

The fourth stage is where the architect Agent writes the technical solution, which includes how to split development tasks and hand them over to the development Agents.

The fifth stage is where the front-end and back-end development Agents carry out the specific software development implementation, and the development results can be previewed in real time during the software development process.

The final link of the entire development is the testing Agent, which first generates complete test cases, then executes the tests, then issues a report, and finally pushes to the Git repository, supporting branch management and multi-person collaboration, and can be uploaded to the environment provided by the platform.

For existing systems, before these processes, there is a stage called the code map. Enterprises can import source code or Git links, and the Agent will parse the front-end, back-end, and configuration files, organize the original logical relationships into charts, and then develop according to requirements. This is mainly applied to the upgrade and iteration of old systems.

But these two stages are in fact also very important. The value of industry experts in these two stages is to translate requirements into specifications and carefully verify the artifacts generated by AI. In an interview, Wang Tong mentioned that AI cannot take the blame for people, that is, deliverables cannot be directly thrown to the business party and need human verification.

At the organizational level, changes have already occurred within NetEase as a result, with the proportion of BA and product roles increasing, and some product development members shifting to positions that better understand the business and acceptance.

In terms of security, Wang Tong introduced that CodeWave SDD mainly provides protection from three levels:

First, it supports local deployment of the entire Agent team and enterprise Harness platform, facilitating integration with enterprise code repositories, knowledge bases, and tool libraries;

Second, a security vulnerability detection system is built into the coding and testing stages, and outputs can only be handed to the next Agent after passing detection. Enterprises can also supplement personalized security requirements such as 4A certification;

Third, it adopts stepwise invocation, splitting complex requirements into multiple tasks for batch processing, avoiding the one-time upload of complete core assets, and at the same time combining different models according to the responsibilities of different Agents to reduce potential risks.

3. Clear advantages over IDEs, 90% first-pass rate for P0 test cases

In the on-site demonstration, Wang Tong showed the actual development comparison between CodeWave SDD and mainstream IDE-type tools.

The evaluation used the core business system development of a certain provincial-level guarantee company, with a PRD exceeding 100 pages. With the same requirements and model, during development, CodeWave SDD essentially reproduced the functional modules, while the IDE only produced the first-level menu, with all submenus lost.

Looking at page details, on the ledger overview page, CodeWave SDD's tabs dividing metric cards and BI charts were both built according to the requirements. The IDE, however, had various problems, such as not dividing tabs by perspective, and missing bar charts and dual-color line charts, among others.

In response, Wang Tong suggested that the reason for this situation was that the contextual requirements exceeded the IDE's scope of handling, and some complex tasks must be broken down by dedicated personnel before being provided to IDE-type tools, so it is not automated development.

CodeWave SDD, by contrast, can accept vague requirements and then design and refine the requirements based on subsequent questions, achieving so-called autonomous driving.

NetEase disclosed that CodeWave SDD's first-pass rate for P0-level test cases is 90%, and about 80% for P1 and P2. However, this metric reflects the pass situation of the first execution of test cases and is not equivalent to the delivery success rate of the entire project.

4. Not just products, but also services: three steps to avoid "shit mountain code"

Previously, in an interview with Zhidongxi, Wang Tong mentioned that in addition to the product itself, CodeWave SDD also provides services such as model recommendation, local computing power support, customer success, and talent training.

In terms of model selection, each Agent can be configured with an independent model. NetEase will conduct full-scenario evaluations after new models are released and form a report within 24 hours to recommend suitable models for different Agents.

For customers such as state-owned enterprises and banks that have local computing power needs, the product also supports local deployment and provides model selection and configuration suggestions.

In terms of services, NetEase will provide enterprises with organizational transformation and talent training services. NetEase will train developers to sort out requirements and gatekeep acceptance, helping them shift from purely technical execution to roles that understand the business and communicate with customers.

Wang Tong believes that an FDE (Forward Deployed Engineer) is not just capable of building Agents; more importantly, they can collect scattered business requirements at the customer site, combine business understanding to sort out the real requirements, then hand them to AI for execution, and accept and modify the output.

Faced with Zhidongxi's question about the "shit mountain code" problem that AI-generated code may be difficult to maintain, Wang Tong used three sentences to explain CodeWave SDD's solution in this regard.

First, use fine-grained Specs as a unified standard. Requirement specifications need to be sufficiently clear, and all Agents work based on this, ensuring that requirements, development, and testing share a common judgment standard.

Second, conduct mandatory verification through the Harness environment. From prototype to code, the output of each stage must be checked for compliance with the Spec. If it fails the check, it cannot proceed to the next stage.

Third, split large tasks into small tasks and execute them in batches. A complex requirement can be split into hundreds of subdivided tasks, with only part of them executed each time. After completion, check against the Spec first, then continue with the next batch, reducing deviations caused by overly large tasks.

Wang Tong said that NetEase hopes that through these three methods, the code generated by Agents will better conform to specifications and be easier to maintain and iterate subsequently.

NetEase's launch this time targets the enterprise AI software delivery market, but unlike conventional AI Coding tools that let developers produce code faster, what NetEase is doing is letting Agents take on more work originally done by humans.

The automation that CodeWave SDD wants to achieve is to let Agents occupy part of the ecological positions of employees and achieve an increase in the number of employees, thereby affecting production capacity. Whether this form can work depends not only on the underlying model's capabilities, but is also closely related to requirement processing, delivery quality, and maintenance capabilities.

According to NetEase's vision, enterprises can increase the number of Agents to take on more development work, allowing people to shift their energy to requirement sorting, architecture review, and delivery acceptance, thereby changing team division of labor and the production capacity structure.

However, software development involves complex business logic and long-term maintenance. From the actual situation, it is still unrealistic to achieve Agent "autonomous driving," and humans still need to strictly gatekeep requirements and acceptance.