How a Master‑Child AI Agent Took Over a Spring Boot Project and Transformed the Development Team

I began a risky experiment: using Codex together with a "master‑child" AI agent architecture to directly take over a real Spring Boot project. My initial question was whether AI could truly participate in software engineering beyond simple code completion.

The AI quickly proved that traditional AI coding—auto‑completing code, generating functions, CRUD, or SQL—was insufficient for a complex project that includes multiple Maven modules, MyBatis XML, Redis, MQ, Feign, gateways, Spring Security, and legacy code. The project’s true nature is a multi‑module, dependency‑rich system.

The master agent first scanned the entire codebase, identifying Maven modules, package structures, services such as OrderService and PaymentService, MQ consumers, Redis keys, and database schemas. It then began reasoning about the impact of a new requirement: “add refund capability to the order system.” Without being told which classes, tables, or interfaces to modify, the agent inferred the affected modules.

At this point the AI stopped writing code and started “understanding the engineering context.” It broke the high‑level requirement into concrete tasks and assigned them to four specialized sub‑agents:

Database Agent : added refund_status and refund_time columns to the order table, updated the corresponding Entity and MyBatis Mapper XML, and even adjusted database indexes.

Service Agent : introduced refund transaction logic, added Redis idempotency handling, created new MQ messages, and modified the order status flow.

API Agent : updated the Controller, DTOs, added parameter validation, and synchronized the Swagger documentation, effectively acting like an API engineer.

Debug Agent : focused on compilation, testing, and error correction. When a sub‑agent caused the project to fail to start, the master agent read the stack trace, identified a bean‑related issue, invoked the Debug Agent, which automatically fixed the problem and reran the project.

This orchestration resembled a real tech‑lead managing a team: the master agent coordinated workflow, managed context, and scheduled sub‑agents. The experience convinced me that AI is no longer just a tool that writes snippets; it is beginning to act as a participant in the full development lifecycle.

Why is Java especially suitable for this master‑child approach? Spring Boot projects naturally adopt a multi‑module, layered architecture—api, service, domain, infrastructure, gateway, mq, common, and often further split into auth, payment, order, user, inventory, etc. Such modularity maps cleanly to distinct agent responsibilities, making task delegation straightforward.

Consequently, the most dangerous shift is no longer whether AI can write code, but whether AI will start “organizing the development process.” With long‑context memory, repo‑wide awareness, tool‑calling, and terminal automation, AI agents can now perform engineering tasks that were previously pure manual labor. This signals a future where AI handles the heavy lifting of software development while developers focus on governing and guiding the AI.

In summary, the experiment shows that AI agents can understand project structure, decompose features, modify code across layers, debug failures, and manage workflow—effectively becoming a simulated development team within a Java Spring Boot ecosystem.