Agents
Cursor Agents Window: Multi-Agent UI, Cloud Runs & PR Handoff
The Agents Window is Cursor's agent-first interface: run parallel cloud agents across workspaces, review diffs and PRs in one place, hand off with /in-cloud and /babysit, and use worktrees for isolated Git checkouts. Open it with ⌘⇧P → Open Agents Window; return to the classic editor with Open Editor Window. Enterprise admins can gate rollout for the first two weeks after Cursor 3 launch.
On this page
- How do I switch between Agents Window and the editor?
- What are the panes in the Agents Window?
- How do I commit to a different branch from the review pane?
- Can the Agents Window drive IDEs Cursor can't natively run in?
- What can I do only in the Agents Window?
- Foreground or background: which agent runs the task?
- When should I stay in the classic editor instead?
- How do Enterprise admins control Agents Window access?
How do I switch between Agents Window and the editor?
You move between the two surfaces from the Command Palette, and the move is reversible at any time.
- 1To Agents WindowCursor's surface listing your agent runs; open a run to read its diff and search transcripts from the command palette. Press Enter for the full definition.: Command Palette → Open Agents Window (⌘⇧P), or the Agents Window button top-right of the editor.
- 2Back to editor: Command Palette → Open Editor Window, or the small arrow in the top-right.
- 3From scratch: File → New Agents WindowCursor's surface listing your agent runs; open a run to read its diff and search transcripts from the command palette. Press Enter for the full definition. opens a fresh agent-first instance.
- 4Quick file peek: ⌘P (file search) and ⌘⇧F (search all files) work inside Agents WindowCursor's surface listing your agent runs; open a run to read its diff and search transcripts from the command palette. Press Enter for the full definition. without leaving agent mode.
The Agents WindowCursor's surface listing your agent runs; open a run to read its diff and search transcripts from the command palette. Press Enter for the full definition. is a ground-up native app, not the VS Code fork that powers the classic editor. The shift toward it is deliberate. As one Cursor field engineer put it, the company has moved away from the file-tree-and-debugger view toward putting the agent front and center, which matters most for designers, PMs and other non-engineers who never wanted the IDE surface in the first place.
Over the past couple months we've seen a departure from this type of view and we've seen more of a focus on agents.
This is covered hands-on in Agent Mode Foundations — 6 short modules, free to read.
What are the panes in the Agents Window?
The layout has three regions. A central prompt window kicks off agents. The top-left agent pane lets you open multiple workspaces or repositories in one instance and run agents across all of them, each workspace carrying its own co-located indexing. The right-side panel is for traditional inspection.
- File browser
- Navigate the workspace like the classic explorer.
- Sandbox terminal
- A per-agent terminal so each run's shell stays isolated.
- Cursor browser
- Drive and screenshot a running app for verification.
- Review pane
- Inspect every edit an agent made before you commit.
Each workspace you open will contain its own indexing. So this kind of allows you to run multiple agents across multiple projects, switch between the progress of them, and see them unfold in real time.
Cloud agents surface here too: you can see and kick them off from the same window. One concrete adjacent win: SSH through the Agents WindowCursor's surface listing your agent runs; open a run to read its diff and search transcripts from the command palette. Press Enter for the full definition. has been updated to be more functional for remote development than the IDE, which helps when your target is a remote or embedded platform reached over SSH.
How do I commit to a different branch from the review pane?
Agents can review their own work in the review pane, and you can branch and commit straight from it. The current branch shows top-right and on the review tab (top-left once there are changes). Next to the commit-and-push button is a small arrow.
If you made changes on main but want them on a different branch, click that arrow button to the right of commit and push. It creates a new branch and commits there instead.
Can the Agents Window drive IDEs Cursor can't natively run in?
Yes, by running it as a sidecar. For mobile work the standard move is the Agents WindowCursor's surface listing your agent runs; open a run to read its diff and search transcripts from the command palette. Press Enter for the full definition. open right next to Xcode or Android Studio, jumping back and forth between them. The Agents Window bridges the gap to IDEs Cursor can't drive natively, with the flow running both directions.
The approach we're going to take for mobile development is using that agent window as a sidecar to help enhance both iOS and Android IDEs and have the integration flow between the two in a bidirectional way.
That bridge runs on ACPAgent Client Protocol. An open standard that lets any IDE host any coding agent; Cursor uses it to run inside JetBrains/IntelliJ and Android Studio. Press Enter for the full definition. (the Agent Client ProtocolAn open standard that lets any IDE host any coding agent; Cursor uses it to run inside JetBrains/IntelliJ and Android Studio via the Cursor CLI bridge. Press Enter for the full definition.), and it reaches well beyond Android. Neovim and Emacs have open-source ACP integration work, and ACP shows up in non-coding tools too. In Obsidian, for instance, you can have Cursor act as an agent inside your note editor while you write. As the spec evolves, Cursor stays hand-in-hand with it.
What can I do only in the Agents Window?
The Agents WindowCursor's surface listing your agent runs; open a run to read its diff and search transcripts from the command palette. Press Enter for the full definition. is built for orchestrating many agents at once, so it carries capabilities the editor does not. You reach for these once agents write most of the code and your job shifts to managing the runs. It is also the home base for non-coding knowledge work: it reframes Cursor as a platform for building with AI agents, giving you the most capable models plus the harness that turns plain language into agent-actionable output, rather than just an editor.
Running many agents at once changes the role. A Cursor support engineer described spinning up one agent per incoming ticket ("how do I log in with SCIMSystem for Cross-domain Identity Management. A standard for automatically creating and removing user accounts when people join or leave. Press Enter for the full definition.", "what does the analytics dashboard show"), reviewing the work as it streams in and tapping into each individually with little context switching. Time is the scarcest currency in support, and this lets one person keep a steady flow of ticket volume without serializing on a single agent.
- Multi-workspace: manage agents across projects from one surface.
- Parallel cloud agents: run many agents in the cloud; monitor from desktop, web, Slack, GitHub or Linear.
- Diffs and PRs: review, commit and manage pull requests without switching apps.
- Cloud subagents:
/in-cloudand/babysitfor long-running VM work on its own branch. - Worktrees: isolated Git checkouts so parallel tasks do not stomp the same files.
Foreground or background: which agent runs the task?
Coding with these tools moved through three eras. First Tab predicted your next edit while you typed. Then synchronous agents edited across files while you watched. Now asynchronous cloud runs go off and finish a goal on their own branch. The split that matters day to day is the last one: keep a task in the foreground when it needs your judgment, send it to the background when the output proves itself.
Interactive widget. Tab through its controls; the result updates in the panel below as you change them.
If the result is checkable through an artifact, it can run async; if the task is ambiguous, keep it in the foreground.
Foreground is for decisions. The task is ambiguous, it needs steering, you want iterative front-end feedback, you are monitoring it live, or it touches something security-sensitive. You see and control every step. Background is for execution against a clear goal with acceptance criteria, where the output is artifact-verifiable: an image, a recording, a log, a diff you can read after the fact.
If the output can be reviewed through an artifact, the task can run async in the background. If it stays ambiguous, keep it in the foreground where you can intervene.
- Task
- Explore an unfamiliar API or spike a feature
- Where it runs
- Foreground
- Why
- The path is unknown; you steer as the agent learns the shape.
- Task
- Live-observed UI tweaks
- Where it runs
- Foreground
- Why
- You watch the rendered result and correct mid-flight.
- Task
- Refactor with a passing test suite
- Where it runs
- Background
- Why
- The diff plus a green run is the artifact you review.
- Task
- Branch migration across many files
- Where it runs
- Background
- Why
- A clear goal and the resulting diff make it verifiable async.
- Task
- Database migration with a checked plan
- Where it runs
- Background
- Why
- Acceptance criteria are explicit; you read the log and diff.
| Task | Where it runs | Why |
|---|---|---|
| Explore an unfamiliar API or spike a feature | Foreground | The path is unknown; you steer as the agent learns the shape. |
| Live-observed UI tweaks | Foreground | You watch the rendered result and correct mid-flight. |
| Refactor with a passing test suite | Background | The diff plus a green run is the artifact you review. |
| Branch migration across many files | Background | A clear goal and the resulting diff make it verifiable async. |
| Database migration with a checked plan | Background | Acceptance criteria are explicit; you read the log and diff. |
Same agent, different home: ambiguity keeps it local, a checkable artifact sends it to the cloud.
Cursor reports that roughly 40% of its own merged PRs are now opened end-to-end by cloud agents. That is a vendor figure, so read it as dogfooding rather than an industry benchmark. Even scoped that way, it shows asynchronous runs against clear goals already carry a real share of shipped work.
When should I stay in the classic editor instead?
Use the editor when you want VS Code extensions, flexible split panes and manual file navigation across many open tabs. Use the Agents WindowCursor's surface listing your agent runs; open a run to read its diff and search transcripts from the command palette. Press Enter for the full definition. when agents write most of the code and you orchestrate parallel runs. You can switch anytime; Cursor supports both experiences.
How do Enterprise admins control Agents Window access?
For two weeks after the Cursor 3Cursor's agent-forward interface (also called the agent window or Glass), built to run and supervise many agents at once rather than edit one file. Press Enter for the full definition. launch window documented by Cursor, Enterprise admins could restrict Agents WindowCursor's surface listing your agent runs; open a run to read its diff and search transcripts from the command palette. Press Enter for the full definition. to specific users or enable it org-wide from Team settings. After that rollout period, access is on by default unless your org policy says otherwise. Confirm current controls in your admin dashboard.
Frequently asked questions
Can I run local and cloud agents in the same window?
Yes. Agents Window is built for handoff: move a task to the cloud to keep working, then pull it local to iterate quickly.
Do worktrees replace Git branches?
Worktrees are separate checkouts for agent isolation. You still merge through normal Git/PR flows when the task is done.
Is Design Mode available in Agents Window?
Yes. Design Mode runs in the in-window browser; toggle with ⌘⇧D. See the Design Mode guide for visual editing.
Sources & last verified
Cursor ships frequently. Last updated June 24, 2026.
Keep reading
Rather do it than read about it? Run 11 interactive Cursor walkthroughs in a simulated editor. Free, no account needed.