Most teams treat a legacy service as a black box they are afraid to touch. This talk is the opposite: a two-year-old Java/Spring Boot microservice, running in production across three ends — mobile app, vehicle, and cloud — with no documentation at all and zero tolerance for downtime, taken apart and rebuilt with AI agents.
The original Chinese title was 技术债的尽头是 AI：一个微服务从零文档到知识库的重构之路. The English version keeps the part that matters: the debt was never the code. It was the missing knowledge.
What the talk covers#
- How to build system comprehension in a codebase nobody documented, using AI agents as readers before using them as writers.
- A 13-step flow borrowed from Kiro’s design philosophy — understand, define, rebuild, deploy — with the actual git commits and knowledge-base layout shown on screen.
- The “athlete and referee” pattern: separate agents for producing changes and for judging them, so the model that writes the code never grades its own homework.
- Which steps should stay manual. The honest answer is not “let the agent do everything.”
The numbers, as presented#
| Measure | Planned | Actual |
|---|---|---|
| Coding | 1 week | 4 hours |
| Integration testing | 1 week | 2 days |
| First integration pass | — | 20 minutes, 90% of features verified |
| Bugs found | — | 2, both human error |
| Pitfalls documented | — | 5 |
The talk deliberately keeps the failures in. A report that only shows the wins is marketing, not engineering.
The conclusion I actually wanted to land#
Process, not the model, is where the safety comes from. The 13 steps would work with Kiro, with Qoder, or with a plain chat window. The harness is the asset; the model is swappable. That is also why the same flow is reusable on the next legacy service instead of being a one-off story.
Photos from the session#
Three frames, three distances: the room, the talk, and the table.
Slides and recording#
Slides are being cleaned up for publication. If you want the deck or the 13-step checklist before then, email me at pinxiong.cn@gmail.com and I will send it over.
The written version of this talk is a series on the blog: start with AI Coding in practice: rebuilding a zero-docs microservice.