After Two Days Refactoring a Spring Boot Project with Codex, I Rethink Java Development

I spent two days letting Codex directly manage a Spring Boot application instead of merely chatting or generating isolated snippets. The AI performed genuine engineering tasks: changing code, running tests, fixing bugs, adjusting interfaces, and refactoring modules.

Key actions included splitting an order module into domain, application, and infrastructure layers; moving packages and imports; fixing circular dependencies; updating DTOs and XML mappings; and even modifying inter‑module dependencies. When I asked Codex to add idempotent handling to the payment module, it automatically located the payment service, the order entry point, the Redis key, the transaction logic, and the MQ retry code, then suggested where duplicate submissions might occur.

In a separate experiment I instructed Codex to replace session‑based login with JWT. The AI rewrote Spring Security configuration, added a JWT filter, updated the login API, changed Redis token handling, adjusted interceptors, and fixed Swagger definitions. It also corrected compilation errors without my intervention.

These experiences showed a shift from AI as a static code‑completion tool to a dynamic "agent loop" that reads the entire repository, generates code, runs it, encounters errors, fixes them, and iterates. This ability to understand the full project context marks a new stage where AI participates in engineering evolution rather than merely supplementing it.

The implications for Java development are profound. Traditional Java projects involve massive repetitive engineering labor—hundreds of beans, dozens of Maven modules, extensive XML and MyBatis configurations. AI excels at pattern learning and can automate much of this mechanical work. Consequently, the most valuable Java engineers will likely be those who can orchestrate AI, design system architecture, and manage AI‑driven workflows, rather than simply writing CRUD code.

Spring Boot’s multi‑layered, modular architecture, strong typing, and strict conventions make it especially amenable to AI assistance. As capabilities like long‑context repository understanding, tool‑calling, and terminal automation mature, AI agents will increasingly become genuine participants in the software development lifecycle.