Junie or AI Assistant? JetBrains ships both for a reason
Open a current JetBrains IDE and you'll find two AI products side by side: AI Assistant and Junie. Teams keep asking which one is "the real one", as if one were a beta of the other. Wrong frame. They're different tools for different postures of work, and shipping both is one of the more honest product decisions in this space.
AI Assistant is the interactive layer: chat about code, completion as you type, explain-this-class, quick fixes, a test generated for the function on your screen. You stay in the driver's seat; the tool shortens your reach. Junie is the delegation layer: you describe an outcome, it plans, edits across files, runs tests, and returns a changeset. That's an agent in the full sense: you leave the driver's seat and become the reviewer.
Two postures, one table
| Dimension | AI Assistant | Junie |
|---|---|---|
| Who drives | You | The agent |
| Unit of work | A question or completion | A task with a definition of done |
| Latency | Seconds | Minutes |
| Your role | Author | Reviewer |
| Cost of a miss | A suggestion you ignore | A changeset you must read and reject |
Triage is the skill
The real question is which tool the task belongs to, and the routing rules are learnable:
- One file, one thought: Assistant. Renames, explanations, a tricky expression, a unit test for what's on screen. Interactive latency keeps you in flow; an agent here is a slow detour.
- Cross-cutting and specifiable: Junie. Thread a new field through DTO, service, storage and tests; migrate a deprecated API across a module. If the acceptance criteria fit in three sentences, delegate.
- Fuzzy: Assistant first, then Junie. Explore in chat until the requirement sharpens, then hand the sharpened task to the agent. Exploration is cheap in chat and expensive in agent-minutes.
Assistants shorten your reach while you keep driving; agents take the wheel and hand you the review. Route per task, not per religion.
Shared quota makes triage real
This would stay philosophical except for one detail: both products draw on the same AI quota under JetBrains' plans, as of this writing. An agent run consumes far more model usage than a chat exchange, because planning, file reads, edits, and test output all cycle through the model repeatedly, so a chat-sized question sent to Junie burns delegation budget for nothing. Running out of quota mid-month is usually a triage failure before it's a capacity problem; Junie's quota economics breaks that down.
The boundary will keep drifting. Assistants industry-wide keep absorbing light agentic tricks, and Junie handles trivial tasks fine, just wastefully. Triage errors cut both ways, too: send a cross-cutting refactor to chat and you become the copy-paste executor of a plan nobody verified. So run the deliberate version for a week (everything interactive to the Assistant, one well-specified delegation a day to Junie) and audit the failures on Friday. Most will be routing errors, not model errors. When you want the view beyond JetBrains, the cross-tool comparison is the map.