Air is a control room, not an agent.
JetBrains opened the Early Access Program for Air in its IDEs on October 1. The plugin can connect Codex, Gemini, GitHub Copilot, Claude Agent, Junie, and other agents that support the Agent Client Protocol. JetBrains says Air ships with no agent installed and is not an AI provider.[1][2]
That distinction is useful. The IDE can show sessions, changed files, outgoing commits, diffs, and cost. The agent still executes the task. The provider still receives the requests covered by its own agreement. A local or cloud worker still owns the process and files. Air can make those routes easier to inspect, but the window does not merge them into one trust boundary.
A unified screen is not unified custody.
Parallel sessions move the hard part.
Chat puts one conversation in front of you. Air is designed to track several sessions across projects. JetBrains also offers temporary worktrees so sessions can start on separate branches or detached work, then return through a cherry-pick.[1]
That is the right mechanical shape. The review problem grows with every active session. One session can be waiting for an answer while another changes a database migration and a third spends against a different provider account. The operator needs a short session manifest before those lanes blur together.
ACP connects the editor and agent.
The Agent Client Protocol standardizes communication between code editors and coding agents. Local agents can run as editor subprocesses over JSON-RPC on standard input and output. Remote support uses HTTP or WebSocket, but the ACP documentation says full remote-agent support is still in progress. ACP also reuses Model Context Protocol representations where practical and adds coding-interface types such as diffs.[3]
Protocol support reduces custom integration work. It does not choose your approval policy, cap spend, isolate branches, or decide when a patch is safe to land. Those remain operating rules.
Make each session carry a tag.
Before dispatch, record five things: the task, agent and provider, execution location, branch or worktree, and the person who accepts the result. Add a spend stop if the route can bill by token or time. Save the outgoing revision and incoming commits. Review the diff in the IDE, then run the checks that match the changed system.
JetBrains already exposes project knowledge to agents through IDE skills in Rider. Its 2026.2 release documents skills for profiling, test placement, code-quality hooks, and framework work.[4] That can make the agent better informed. It does not turn an inspection or a profiler hint into a passing build.