AI code generation in Java
How generative AI has transformed software development
Who am I?
- Principal Manager, Java Developer Relations @ Microsoft / GitHub
- Creator of JHipster — Open Source platform with 22,000+ ⭐
- Contributor to LangChain4j & Spring AI
- 200+ international talks (Devoxx, SpringOne, Microsoft Build…)
- Author of "Spring par la pratique" — the first French book on Spring
- 25+ years of experience in the Java ecosystem
The evolution of AI agents
Agent Mode, Plan Mode & Git Worktrees
🤖 Agent Mode
The AI acts autonomously:
- Reads project files
- Writes and edits code
- Runs shell commands
- Executes tests automatically
- Fixes errors on its own
📋 Plan Mode
The AI plans before acting:
- Analyzes the request in depth
- Generates a detailed action plan
- Asks for your approval
- Executes step by step
- Backtracks if needed
🌳 Git Worktrees
Multiple agents in parallel, with no conflicts:
- One directory per branch
- One agent per worktree
- Several tasks at once
- Full isolation of changes
- Controlled final merge
git worktree add + 1 agent per folder
Model Context Protocol (MCP)
🔌 An open standard protocol
MCP lets AI agents connect to external tools: databases, APIs, file systems, cloud services — in real time.
Demo: Using GitHub's MCP server
Use the GitHub official MCP server to find: - In repositories coded in Java, with more than 100 stars - 10 tickets with the "good first issue" label - Tickets should have a title, GitHub repository, link
Skills
🎯 What is a Skill?
A skill lets you specialize an AI agent
with precise domain instructions,
context files and dedicated tools — for a specific domain.
Open standard Agent Skills
(agentskills.io)
— compatible with Copilot CLI, Claude Code, Windsurf…
📦 dr-jskill
My skill for building Spring Boot 4.x applications with Java 25, PostgreSQL, Docker, and your choice of a Vue / React / Angular front-end.
github.com/jdubois/dr-jskill💡 Why use skills?
- Custom coding conventions
- Architecture defined upfront
- Reproducible results
- Shareable across the team
Demo: Spring Boot 4 + Java 25 + Vue.js with dr-jskill
Create a ticket management application with Dr JSkill.
Here are the specifications:
- the application's goal is to find good tickets to help Java Open Source projects
- tickets are GitHub tickets, as discovered with the MCP prompt from the previous demo
- add/edit/remove tickets with a title, GitHub repository, link and status
- store them in a database, and initialize the database with the tickets from the previous command
- it has a fancy UI with Vue.js
IDE Integration
Java IDEs now integrate AI natively.
🟦 VS Code
- Native GitHub Copilot — built by Microsoft, the deepest integration
- Agent mode — completion, chat and agent mode
- Java Extension Pack — JDTLS, debug, tests, Maven/Gradle — used by the agent
- Choose your model (GPT, Claude, …) within Copilot
🟥 IntelliJ IDEA
- GitHub Copilot plugin — completion, chat and agent mode
- ACP plugin (Agent Client Protocol, an open standard launched by Zed) — plug an external agent (GitHub Copilot CLI…) directly into the IDE
- IntelliJ MCP server — exposes the PSI index, refactorings and inspections to external agents
- JetBrains Junie — JetBrains' own agent
- Access IntelliJ's semantic refactorings (rename, extract, change signature…) from the agent
AGENTS.md — the project brief
A single file at the root, read by all agents: the open AGENTS.md standard.
🧭 What it contains
- Project overview, stack, ports
- Exact commands:
./mvnw verify,./mvnw spring-boot:run - JDK version, framework, BOM
- Code conventions, tests, coverage
- Commit & PR conventions
- Forbidden actions (prod, secrets,
--force…)
🚀 How to create it
/initin Copilot CLI — the agent reads your repo and proposes a draft- Review and trim: short beats long (context rot)
- Iterate: every time the agent gets it wrong, fix
AGENTS.md— not the prompt - Versioned, reviewed in PRs just like code
🗂️ Monorepo & modules
AGENTS.mdat the root = global rules- One
AGENTS.mdper sub-module for specifics (closest wins) - Point to expected skills and MCP servers
- Link to
docs/rather than duplicating everything
AGENTS.md, the agent rediscovers the project every session — and makes the same mistakes in the same places
copilot-instructions.md with /init
Demo: Generating a copilot-instructions.md with /init
Best practices for Java
Modern agentic CLIs (Copilot CLI, Claude Code…) integrate the LSP protocol. For Java, that means JDTLS (Eclipse JDT Language Server).
☕ JDTLS — Eclipse JDT Language Server
- The same engine as VS Code Java & Eclipse, exposed via LSP
- Full semantic analysis: AST, type system, Maven/Gradle classpath
- Safe refactorings: rename, extract method, change signature — propagated across the whole project
- Navigation: go-to-definition, find-references, type hierarchy
- Quick-fix & organize imports, driven by the agent
- Real-time diagnostics: compilation errors, warnings, nullability
✅ Project hygiene
- Pin the JDK
./mvnw compilemust be fast- Keep
maingreen — it's the agent's source of truth - Loop: generate →
verify→ read the errors → fix
🚫 Not recommended inside an agent loop
-DskipTests — tests must run.
Harness > model
An agent = tools in a loop. At equal model quality, it's the harness around the model that makes the difference between an agent that ships and one that hallucinates.
🔁 The loop
- Observe → decide → act → verify, continuously
- Plan mode before agent mode on any non-trivial task
- Sub-agents: isolate contexts, parallelize
🛠️ The tools
- Skills, MCP servers, LSP, sandbox shell — deterministic capabilities
- Beyond ~20 loaded tools, the model picks the wrong one
🧠 The context
AGENTS.md+ skills: brief the agent once and for all- Avoid context rot: poisoning, distraction, confusion, clash
- Context offloading:
plan.md, memory, summaries — not everything in the window
✅ The verdict (evals)
- A machine-verifiable criterion:
./mvnw verify - The agent reads the errors (LSP, tests) and fixes them on its own
- Without a clear verdict, the agent spins — and no model will save it
💡 Practical consequence
The same model, in two different harnesses, produces very different results (see SWE-bench rankings). Investing in the harness — skills, MCP, AGENTS.md, LSP, verification loop — pays off more than chasing the latest model.
GitHub Copilot App
Technical preview — a native desktop app from GitHub: start an agentic session, isolate it, steer it, and merge the final PR.
🧭 Start from GitHub context
- Kick off from an issue, PR, prompt or a previous session
- Unified inbox across all connected repos
- Details, reviews and checks stay attached to the session
⚡ Isolated sessions
- Each session: its own git worktree, files, conversation
- Pause / resume, in parallel across multiple repos
✅ Steer, validate, ship
- Review the plan and the diff, guide iteration
- Validate via the built-in terminal and browser
- Open the PR — same reviews, same checks
- Agent Merge: reviews, checks, automatic merge
Demo: 3 parallel sessions with the GitHub Copilot App
I want to assign tickets to users. There are 3 users, stored in a database table: julien, alice, bob. Update the database, the Java code and the front-end code accordingly.
Conclusion
🚀 What has changed
- From completion to autonomy (agent + plan mode)
- The importance of the harness: MCP & Skills, LSP, AGENTS.md
- GitHub Copilot App & Git worktrees = agents in parallel
- Deep integration in VS Code & IntelliJ
✅ What stays true
- The developer stays in charge of the code
- Technical understanding is key
- JHipster is still relevant! 😄