JetBrains Air: an IDE built around the agent, on Fleet's bones
I've run IntelliJ as my daily driver since roughly 2011, so when JetBrains shipped a brand-new IDE this spring I approached it the way you approach a friend's rebound: politely, and braced for it to end badly. Air arrived as a macOS-only public preview in March, and I ignored it for a couple of months. Late June I finally moved a slice of real client work into it: a Kotlin service in a logistics backend I've maintained for two years. I gave it an honest week.
The pitch fits on a napkin. Instead of bolting a chat panel onto an existing editor, JetBrains built the tools around the agent from the start, so the agent is the primary user of the workspace and you're the reviewer. That reads like a slogan until you sit in it for a few days, and then the shift in posture gets hard to unsee.
the part that won me over
What sold me was small and unglamorous. The unit of work in Air is a task, not a file. I describe what I want, an agent does it inside an isolated workspace, and I come back to a diff I approve or reject. I cover the mechanics in Air's task model, so I won't rehash them, but the effect is that the whole surface is arranged around that loop instead of around a tree of open buffers.
By day three I'd stopped opening a terminal tab out of reflex.
Fleet's ghost
Now the part the press couldn't stop prodding. Air is built on the codebase of Fleet, JetBrains' previous next-generation editor (the one that spent years in preview and never grew up). Building your new flagship on the bones of the abandoned one earns every bit of the skepticism you'd expect, and I felt it too. My first afternoon I kept bracing for the same vague, floaty feeling Fleet always gave me.
It mostly isn't there. The chrome is quieter, and the agent loop gives the whole thing a reason to exist that Fleet never had. But the lineage shows in odd corners: a settings panel that looks half-drawn, a keybinding that silently does nothing, the occasional half-second lag that feels like an old memory surfacing.
Fleet was an editor hunting for a reason to exist; Air is that reason, arriving about three years late.
the week I actually spent in it
My honest blunder that week: I ran Air with the wrong agent for two days and blamed the IDE for the results. Air is agent-agnostic. Out of the box it talks to OpenAI Codex, Claude Agent, Gemini CLI, and JetBrains' own Junie, so the brain doing the work is a setting, not a fixed part of the product. I'd left it pointed at a model that's fine for chat and mediocre at multi-file edits on a JVM codebase, so my first tasks came back shallow and I nearly wrote the whole thing off. Swapping the agent fixed most of what I'd been grumbling about, which hammered home the obvious point that the environment and the agent are separate things you tune separately. If you're already lost among Air, Junie, and the AI Assistant inside IntelliJ, you're not alone, and I untangle all three in Air, Junie, or plain IntelliJ.
Once the agent was right, the numbers got interesting. On a nasty refactor across our 400k-line monorepo (moving a batch of controllers onto a new validation layer), Air took roughly 70% of the mechanical edits off my plate and left me the 30% that actually needed judgment. Not sorcery, but a real dent in an afternoon.
what still feels unfinished
Two things stop me from telling the whole team to switch. It's macOS-only right now, and half my team lives on Linux; Windows and Linux builds are promised but not here, so Air can't be a team default yet. The other catch stings more: it's not IntelliJ. Fifteen years of muscle memory doesn't carry over: the refactorings, the debugger, a thousand tiny shortcuts. Whenever I need heavy structural surgery I still bounce back to IntelliJ and its agent Junie.
So I run both, Air for agent-driven tasks, IntelliJ for the deep surgical work, and I've made an uneasy peace with the split.
Is Air the future, or a preview that flatters to deceive? Ask me in six months. What I can say is that the build-around-the-agent bet feels correct in a way that stapling chat onto an editor never quite did, and I think the whole category is drifting toward this shape, a hunch I chase down in where agentic IDEs are converging. For one honest week of Kotlin work, Air earned its spot in my dock, which is more than Fleet ever managed.