Junie Tops SWE-Rebench at 61.6%, JetBrains Ships Agentic Debugging and Plan Mode

Junie scored 61.6% resolved and a 72.7% pass@5 on the latest SWE-Rebench cycle, placing number one among coding agents and competitive with raw frontier models. That's the benchmark JetBrains used to declare its AI coding agent ready for general availability.

Plan Mode Cuts Wasted Tokens Before the First Line of Code

Most AI agents charge ahead implementing the wrong thing. Junie's Plan mode forces a structured document first: product requirements, technical design, delivery stages, testing strategy. You edit it in your editor, approve it, then Junie implements. The plan lives in .junie/plans and can be committed as living task documentation. When requirements are ambiguous, Junie asks multiple-choice or freeform questions instead of guessing. You plan on a strong model and implement on a cheap one, keeping bills low.

Another common failure pattern - agents that guess at runtime state - gets the same treatment. Junie drives the IDE's real debugger: launch a debug session, set breakpoints in project code, library code, even decompiled .class files. It inspects stack frames, evaluates expressions, and runs to line. No more println archaeology. You can hand off tasks like "debug why this test fails only on the second iteration" and let Junie work while you focus elsewhere.

Remote Control and Context-Aware Code Review

Tasks that take hours - a Spring Boot upgrade, a migration to Java records - run asynchronously. Start from your laptop, check progress from your phone, review the PR later. Junie keeps the session accessible anywhere you sign in.

Code review gets the same project context Junie uses to write code: your build, your tests, your past decisions. Trigger a review from GitHub Actions, GitLab (including on-prem), or via CLI. Junie highlights each meaningful change, explains design decisions, and lets you accept or reject inline. Drop a PR comment on the spot. Ask a follow-up question and Junie reorders the remaining review around what you care about, instead of marching through files alphabetically.

One Engine, Many Surfaces, Any Model

Junie's GA integration is rebuilt on ACP (Agent Communication Protocol), the same protocol the CLI uses to talk to your IDE. Improvements ship once and show up in the AI chat, the dedicated Junie tool window, and the terminal. The agent uses your IDE's semantic index, build configurations, test runners, and debugger - not its own approximation. Database integration means Junie connects to your DataGrip configured databases, queries real data, and writes, fixes, and validates SQL in the same session as your code.

Cost control is a dial you hold: support any model via BYOK (Anthropic, OpenAI, Google, others) or point Junie at local runtimes like LiteLLM, LMStudio, or Ollama. Prompts and code never leave your machine.

Each feature alone solves a sharp problem. Together they change what an agent is for: a delegate you trust with plans, debugging, review, and data queries - not a fancy autocomplete. That's the bar JetBrains set for leaving beta, and the feedback loop that built every feature here doesn't stop at GA.

Source: Junie: The JetBrains AI Coding Agent Leaves Beta.
Domain: blog.jetbrains.com