AI-assisted Programming: From Inline Completions to Agentic Workflow

The Evolution of AI in Code Editors

In the session led by Anton Arhipov, the focus was on the evolution of AI-assisted programming within JetBrains IntelliJ, highlighting the new agent tool, Junie. Arhipov mapped out the progression from basic inline completions to more complex agentic workflows, demonstrating these advancements live within the IntelliJ environment.

Minimal Assistance: Inline Completions

Anton Arhipov began by introducing the concept of minimal assistance offered by AI in code editors. At this level, AI provides single-line or statement-based code completions. "At this end, you are fully in control—AI just helps you type a bit faster but never gets in the way," Arhipov explained. This assists developers by predicting the next line of code with a local model, ensuring non-intrusive, relevant suggestions. He emphasized JetBrains’ philosophy of avoiding ‘hallucinated’ completions that could disrupt the developer's workflow.

Enhanced Assistance: Cloud Completions

As the session progressed, Arhipov demonstrated the shift to enhanced assistance through cloud completions. This feature leverages remote models to offer more substantial code snippets, taking into account a broader context such as preceding comments or the overall project structure. Arhipov illustrated how simply commenting an intention could yield a complete, correct code snippet. He elaborated on different filtering modes—focused for safer results and balanced for more creative yet riskier outputs—using the Melom completion model.

Advanced Capabilities: In-editor Code Generation

Arhipov then explored more advanced AI capabilities, including in-editor code generation. This feature allows the AI to make changes across files, add necessary imports, and even refactor code based on minimal prompts. Such actions are automatically checked for correctness and relevance, minimizing the need for manual verification. This level of assistance represents a significant step towards automating routine coding tasks, allowing developers to focus on higher-level problem-solving.

Agentic Workflows: The Role of AI as an Orchestrator

The session culminated in a discussion of agentic workflows, where the AI acts as an orchestrator. Arhipov showcased how developers could interact with AI through a chat interface to perform tasks such as rewriting code or generating unit tests. At this stage, the developer becomes an orchestrator, choosing which AI-generated changes to implement and how to validate them. “At this point, we’re at a position of being an agent. The developer is still ultimately responsible for integrating and testing AI-generated code,” Arhipov noted.

Introducing Junie: The Next Generation Agent Tool

Wrapping up the talk, Arhipov introduced Junie, the newly-announced agent tool designed to automate full programming workflows. Unlike previous AI assistants, Junie can orchestrate code changes that span entire projects, from controllers to configuration files. This tool is built primarily in Kotlin, emphasizing integration with various platforms and languages. Arhipov highlighted that despite its advanced capabilities, Junie still requires developer oversight to ensure the correctness and efficacy of its outputs.