跪拜 Guibai
← All articles
Backend · Java · Programmer

IntelliJ IDEA 2026.3 Replaces AI Assistant with JetBrains Air and Ships Java/Kotlin Intelligence to Other Editors

By SimonKing ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

The removal of AI Assistant and arrival of JetBrains Air changes how IntelliJ users run coding agents: Air is a neutral protocol layer, not a bundled AI product, so teams can plug any model without waiting on JetBrains. The LSP bridge also means a Java/Kotlin codebase's semantics can now follow developers into Cursor and VS Code, so editor choice no longer decides which tools understand the codebase.

Summary

JetBrains shipped IntelliJ IDEA 2026.3 on November 5, 2026, and the headline feature is JetBrains Air, a workspace/protocol layer for running coding agents inside the IDE. Air does not come with any agents or models; it just provides the terminal, tooling, and hooks for whatever agent you want to plug in. The old AI Assistant is gone, replaced by Air's dedicated left-sidebar icon and agent terminal. Users can add their own LLM providers, which means no lock-in to a specific vendor's model.

Beyond Air, the release extends Java and Kotlin intelligence to other editors through LSP, so Cursor, VS Code, and Antigravity share the same semantic understanding as the IDE. Debugging gets a Hibernate debugger with SQL, parameters, HQL, JQPL, source navigation, and replay; a Spring cache gutter; and AppGlass Debugger for attaching to running environments. The terminal now uses Ghostty's libghostty, adding modern terminal protocols for Neovim, tmux, and terminal agents. A DuckDB-powered workspace data source joins across different DBMSes. JetBrains also claims 1000+ bug fixes and 200+ performance fixes. The author's take: most upgrades target compatibility with new tech, so everyday developers can skip, but Air is worth trying if you're curious.

Takeaways
— IntelliJ IDEA 2026.3 was released on November 5, 2026.
— JetBrains Air replaces the old AI Assistant; it is not an AI provider and ships without agents or models.
— JetBrains Air provides a terminal for agents and lets users add their own LLM providers.
— Java/Kotlin intelligence now works in VS Code, Cursor, and other editors via LSP.
— The Hibernate debugger adds SQL, parameter, HQL, JQPL support, source navigation, and replay.
— Java 28 EA can be downloaded and configured directly in IDE settings.
— A DuckDB-powered workspace data source enables joins across different DBMSes.
— The terminal now uses Ghostty's libghostty, supporting Neovim, tmux, and terminal agents.
— JetBrains claims 1000+ bug fixes and 200+ performance fixes, including Gradle/Maven sync and JPA inspection freezes.
— The author says most updates are compatibility-focused and not necessary for daily development.
Conclusions

JetBrains Air's protocol-level design leaves model choice to users, making the IDE a neutral host for any coding agent rather than a vehicle for a single AI vendor.

Opening Java/Kotlin semantics to other editors via LSP weakens the lock-in that previously required owning an IntelliJ product to get full language support.

The author's advice to skip most upgrades reflects a common reality: version churn for Java and language tooling is often about compatibility with new ecosystem versions, not features that change daily workflows.

Concepts & terms
JetBrains Air
A workspace/protocol layer in JetBrains IDEs for running coding agents. It does not include models or agents; it only provides the infrastructure, terminal, and hooks for user-supplied agents.
LSP (Language Server Protocol)
A standard protocol that lets editors get language features like completion and navigation from a separate process. JetBrains now exposes its Java/Kotlin intelligence through LSP, enabling other editors to use it.
ACP (Agent Communication Protocol)
The older protocol used by IntelliJ's AI Assistant, which the author found far less capable than other agent plugins. It has been replaced by JetBrains Air in this release.
DuckDB
An in-process analytical database. IDEA uses an in-memory DuckDB instance to power a Workspace data source, allowing joins across data from different database management systems.
From the discussion
Featured comments
星云PLUS团队

Air as a protocol layer ships without bundled agents or bundled models — this effectively turns "who produces the conclusion" into a config item: the same repository and the same prompt yield different patches when you swap the model or the agent, making regression attribution difficult. I'd suggest pinning the agent version in the workspace and keeping a model config snapshot. Also, AppGlass's live tracepoints are genuinely useful for investigating agent behavior — no restart or debug port needed to grab the real input parameters and return values of a given step, which is faster than digging through logs afterward.

SimonKing

Hadn't considered that; I'll give it a try later.

See top comments, translated →
Source: juejin.cn ↗ Google Translate ↗ Backup ↗