While Others Add AI to the Business, Kidswant First Made Its Engineering Team AI Native
When most companies talk about AI transformation, they mean adding a few AI features to the business. Kidswant thought one step further ahead: first make "building software" itself AI native. Its lever is Qoder.
A Full-Stack Requirement Went from One Week to One Day
In the past, delivering a full-stack business requirement meant a person breaking the requirement into tasks and typing code line by line, with a tool offering a few completions on the side. Kidswant's engineering team switched to a different way of working: using Qoder, they described the goals and acceptance criteria clearly and built Zhishu, the company's internal AI platform—cutting the delivery cycle from nearly six months down to two.
Qoder is an agentic coding platform built for real-world software development. Unlike ordinary code-completion tools, it first understands the entire codebase, then advances tasks systematically the way an engineer would. In day-to-day development, several capabilities changed the team's rhythm.
For a system with many modules and tightly interwoven front and back ends, the biggest barrier is "not being able to read this codebase." Qoder's Repo Wiki automatically distills the architecture buried in the code into documentation, keeps it updated as the code changes, and can be committed to Git so the whole team shares one understanding. When a developer asks "how is this part implemented," the answer is there without digging through source code. During the development of Zhishu, multiple engineers collaborated on the platform, and Qoder's Repo Wiki enabled efficient coordination across complex modules—playing a key role especially in the continuous iteration and integration of the Agent orchestration and knowledge base modules.
For actual feature work, the team uses Quest Mode: describe the goals and acceptance criteria clearly, and it first aligns on scope and designs the solution, then writes code end to end, verifies its own work, and fixes its own issues. Full-stack requirements mean touching files across multiple tech stacks at once—exactly where its multi-file editing and long-horizon execution shine. For the Skills Marketplace feature on the Zhishu platform, which involved front-end and back-end interactions, engineers used Quest Mode and finished in about one day after describing the feature requirements and page interactions.
To make AI-written code fit their own business instead of producing boilerplate that needs rework, the team packaged coding conventions, business context, and review standards into project-level Skills, so agents automatically follow them when writing and reviewing code; they then used MCP to connect Qoder to internal tools and systems.
The significance of this shift is not "how much faster the code gets written," but that the team's default way of working moved from "humans write, AI assists" to "humans define, AI delivers." That is what AI Native should look like.
A Dozen Expert Agents, a Dozen New Real Teams for the Company
The most important thing Kidswant built this way is Zhishu, its in-house AI platform. Built on the Qwen large language model, it integrates knowledge bases, workflows, and Skills capabilities, supports multi-agent collaboration and long- and short-term memory context, and works across channels such as DingTalk and H5.
Today Zhishu runs 400+ business scenarios and 300+ agents, and has distilled more than a dozen "expert agents"—covering roles such as data analysis, finance, supply chain, customer service, and content creation—each mapped to a real team inside the company. Notably, most of these skills were also created and refined with Qoder.
The Super Portal is a company-wide unified entry point built on the Zhishu Agent platform (composed of 10+ business expert agents), available from both Kidswant's internal "site" employee workbench and its "Customer-Employee Unity" system.
The same development approach was reused for Zhishu's desktop edition: Kidswant used Qoder to build Mac and Windows clients with 50+ built-in internal operations and office skills, which can also host background automation tasks such as "digital employees."
The results are tangible: financial reconciliation dropped from 3 days of manual work to 20 minutes; data analysis became nearly 3x more efficient; the intelligent settlement assistant resolves more than half of all issues; and the content creation expert saves the company the capacity of thirty-plus people. More and more business staff are now building agents and skills tailored to their own work directly on the platform.
This forms a closed loop: engineers write code with Qoder, the code builds Zhishu, Zhishu serves the business with Qwen, and new business needs flow back to engineering—where Qoder is used again.
Even more importantly, the content marketing expert, supply chain expert, and store operations expert that Kidswant has built are fully adapted to the maternal and child retail industry.
"Can You Use AI" Is Now Part of Performance Reviews
Beyond tools and platforms, Kidswant grounded the transformation in its organization. The CTO line and the HR line led the effort together—upgrading how engineering collaborates while adding "AI proficiency" to capability assessments.
Tools save time, but what you do with the saved time determines how deep the transformation goes. For Kidswant, the development time saved by Qoder became time for the team to build the platform, connect scenarios, and codify methodology.
Final Thoughts
For a retail company, the hardest part is not buying the most advanced tools—it is turning tools into organizational capability. That is exactly what Kidswant is doing.
The path Kidswant has taken comes with several preconditions worth benchmarking against:
● Signals that you should start from engineering: business-side AI requests queuing for more than 3 months, delivery failing to keep up with business demand, and the bottleneck being capacity rather than ideas.
● How to pick the first pilot requirement: do not pick the simplest one (it proves nothing), and do not pick the most complex one (too many variables, easy to fail). Pick a requirement with clear boundaries, multiple tech stacks involved, and roughly one week of effort the traditional way.
● When Repo Wiki delivers the most value: codebases of several hundred thousand lines or more, multi-person collaboration, and newcomers who cannot understand legacy code. A two- or three-person small project may not need it.
● How to decide which Agents to build first: do not start from "what AI can do"—start from "which teams in the company keep doing the same kind of work." Kidswant maps each expert agent to a real team (finance team → finance expert, supply chain team → supply chain expert), so the agent's boundaries, users, and evaluation criteria are all ready-made rather than designed from scratch. Start with teams that are labor-intensive, rule-driven, and produce standardized output.
● When to push it to the organizational level: do not call for company-wide AI adoption from day one (no one believes it without success stories). Get benchmarks first, get numbers first, then roll out the mechanisms.