AI Java workflow repo maps Spring Boot delivery to agent roles
Lanqiao-Luca/ai-java-workflow packages 20 agent role definitions, layered workflow templates and runnable Spring examples, with its own limits written down.
A GitHub repository called Lanqiao-Luca/ai-java-workflow has collected 336 stars and 13 forks in the 13 days since it was created on 2026-11-01, according to its own README on GitHub as of 2026-11-14. It presents itself as an AI Java Spring Boot workflow: documentation, role definitions and small runnable examples for carrying a Java or Spring Boot change from requirement through implementation, testing, review and delivery. The main language listed is Python, the licence is MIT, and the README states that this is not an implementation of, or a tutorial for, the Spring AI framework.
What is actually in the repository
The most concrete part is a set of 20 agent role descriptions. Each one states what the role receives, what it hands back and how the result should be checked. The README groups them by job: moving a requirement to something implementable (requirement owner, solution author, design reviewer, module planner), making and checking Java or Spring changes (implementer, verifier, Java build runner), reviewing a change (code reviewer, semantic reviewer, principle reviewer), diagnosing faults (evidence collector, diagnoser, conclusion reviewer) and coordinating delivery.
There are worked examples too. A Spring Boot batch archive case, which requires JDK 17+ and Maven 3, records that the correct implementation passes all 14 tests while a version with defects deliberately left in triggers 3 expected failures, showing how to check that an invalid entry does not cause a partial write. A second case reports six native sub-agents working in Codex CLI on macOS, ending with 25 offline Spring tests passing. The source is careful to say the design and review portion of one case is document reasoning written by the author, not an independent model run.
How the layers fit together
The architecture is a stack. Standards (rules and acceptance criteria) and tools (execution interfaces) sit side by side, skills combine them, agents use skills, workflows sequence agents, loops repeat a workflow toward one goal under a budget, and graphs order loops and workflows with dependencies and concurrency. A routing component called Brain decides which layer a task should enter.
One detail matters for anyone evaluating the project: the README says Brain is currently a documentation router used by a human or an assistant, and that the existing runners do not read its routing automatically. The CLI also does not authenticate or intercept callers. Higher layers organise lower ones, and a task is not required to use every layer.
Where it runs, and where it stops
Work happens in worktree-isolated directories with an optional sandbox for execution, Docker being one sandbox backend. Four backends are listed with what has actually been exercised: a native agent chain on macOS with Codex CLI and Maven, an experimental macOS sandbox-exec backend that refuses fork and cannot run Maven or Gradle, a macOS trusted build sandbox for Maven and Gradle, and a Docker sandbox tried on macOS controllers and inside a Linux VM. A unified command offers doctor, choose, run, status and resume across five whitelisted cases, and keeps run directories outside the repository.
Setup is light. Clone the repo, then run python3 -B scripts/check_all.py with Git and Python 3.9+. That executes eight offline checks on the public file manifest, document references, case identity and loop control tests. It does not call a model or build a Java project, and it exits non-zero if any check fails.
Limits and open questions
The examples are fictional and deliberately avoid company services or real databases. Production deployment, real authentication, load and compatibility with an arbitrary project are left to whoever adopts it. The internal evaluation is small: across 8 author-labelled tasks, a full contract check scored 8 correct while a simplified field baseline scored 3 correct and missed 5, and the README says that comparison does not measure real model performance, human rework or engineering payoff. Cases recorded as NOT_RUN keep that status. Whether the routing conventions hold outside the whitelisted cases is untested.
What this means for building apps with AI
The transferable idea is not the Java specifics but the decomposition: named roles with defined inputs, outputs and acceptance criteria, plus an explicit separation of what has been run from what has only been written down. Anyone writing instructions for an AI builder can borrow that discipline. Draftship is an AI app builder where you describe an app in plain language and it writes the code, builds and hosts it, with a built-in database, user sign-in and hosting; paid plans add private projects, a choice of AI model, source code access, custom domains and GitHub push. It has no connection to this repository. For prompts that spell out roles, inputs and acceptance criteria, see writing effective prompts.
Primary sources
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.