In this updated talk, Anton from JetBrains presents a comprehensive exploration of AI-assisted programming within IntelliJ IDEA, with a special focus on Junie — a new agentic development tool designed to automate and supervise complex development workflows. Through demos, reflections, and audience Q&A, Anton shows how different levels of AI support — from basic code completion to fully autonomous multi-step agents — can transform the developer experience.

Overview of AI in IntelliJ IDEA

• JetBrains integrates various AI capabilities into IntelliJ IDEA, ranging from lightweight in-editor completions to full agentic automation via Junie.
• Users can choose between local and cloud-based models, including JetBrains-hosted LLMs (like Mellum) and external providers (e.g., Claude 3.7).
• The AI Assistant and Junie aim to augment, not replace, developers — helping them move faster, experiment more freely, and focus on higher-level problems.

AI Support Levels and Capabilities

Anton outlines a progressive scale of AI-assisted development:

Level 0: Minimal Assistance

• Local inline code completions using lightweight models.
• Helps with repetitive tasks when the developer knows exactly what to do.
• Fast and context-aware, but limited in scope.

Level 1: Verbose Completions

• Multiline cloud-based completions based on comments or partial code.
• Useful for solving known problems or generating code in unfamiliar languages.
• Anton shows how it can implement utility functions just from a descriptive comment.

Level 2: Structured Prompting

• Developers use inline prompts or actions to trigger higher-level suggestions.
• Includes tools like “Generate Unit Tests” or “Suggest Refactoring.”
• More control over code generation, with automatic insertion and preview diff support.

Level 3: Conversational Refactoring

• Full-featured chat interface for working with the codebase.
• Supports tasks like changing code style, converting code structure, or adding tests based on discussion.
• Still requires developer supervision — reviewing changes and rerunning tests.

Level 4: Agentic Workflows with Junie

• Junie is an AI-powered agent that can scan your project, generate a development plan, and execute tasks across multiple files.
• Developers can write high-level goals in .junie/guidelines.md and track progress in markdown task lists.
• Example: Anton shows how Junie analyzes a project, writes an improvement plan, breaks it into tasks, and executes “Phase 1” while updating the checklist.

Notable Features and Demos

• Full-Line Completion Plugin: Offline, single-line predictions.
• Mellum Model: JetBrains' own fine-tuned LLM for multi-line completions.
• Automatic Unit Test Generation: Based on context and function analysis.
• Inline Prompting: Use comments like // group list of students by name to trigger completions.
• Chat-Based Refactoring: Claude-powered assistant suggests and applies refactorings with inline explanations.
• Junie:
  • Scans codebase.
  • Generates structured development plans.
  • Executes code changes across the project.
  • Can be guided by persistent markdown instructions.

Anton’s Hobby Projects: AI as a Creative Partner

Anton shares how Junie empowered him to create multiple side projects without writing a single line of code:

• Games (Minesweeper, 2048, Solitaire) in Kotlin Multiplatform.
• Web tools (MIDI visualizer, podcast recorder, music tuner) — all developed without frontend experience.
• AI-driven fitness tracker using pose estimation in the browser.

Practical Insights and Discussion

• Model Selection: Claude 3.7 used for Junie; JetBrains handles API tokens and quota for users.
• Plans and Pricing: Free tier expected with limits; enterprise hosting options also available.
• Gradle Support: Junie understands Gradle configurations and adapts suggestions accordingly.
• Future Features: Planned integration with CI/CD, OpenRewrite, and MCP for deeper automation.

Key Takeaways

• AI tools are most powerful when developers know how to use them strategically.
• Junie exemplifies a shift from AI as assistant to AI as autonomous agent, executing tasks while you supervise.
• Greenfield projects, fast prototyping, and unfamiliar tech stacks are ideal use cases.
• Developer judgment, intuition, and domain knowledge remain essential — but AI tools can dramatically accelerate learning and iteration.