Cursor 3 launches agent workspace for local, SSH, and cloud sessions

• Cursor’s launch thread and the official launch post position Cursor 3 as a separate, agent-first workspace rather than a small IDE refresh.
• According to Cursor’s product thread and the forum announcement, agents can now run across local machines, isolated worktrees, remote SSH targets, and Cursor’s cloud runtime from one sidebar.
• The launch post says sessions can move from cloud to local for manual edits, or from local back to cloud so longer jobs keep running after a laptop closes, which Kevin Kern’s first-look note also highlighted as the core workflow change.
• In the official worktrees post, Cursor turned worktrees into explicit /worktree and /best-of-n commands, with separate checkouts per model and a parent agent that compares outputs.
• After a week of testing, Dan Shipper’s write-up called the rebuild fast and strategically pointed in the right direction, but still early enough that the orchestration layer looks more promising than finished.

You can read the full launch post, skim the worktrees and Best-of-N breakdown, and the product already exposes a dead-simple entrypoint because Eric Zakariasson says the whole thing opens from Cmd+Shift+P → Agents Window. The most interesting product choice is that Cursor did not throw away the old editor surface: AILeaksAndNews’ summary and the official post both frame the new workspace as a companion window that still lets you drop back into the IDE.

Cursor 3’s centerpiece is Agents Window, a new interface that Cursor says it built from scratch around agent workflows rather than around the editor itself. In the launch post, the company describes it as a unified workspace for building software with agents, with a multi-repo layout and a higher-level view of ongoing work.

The official forum announcement adds the concrete operational model:

• run many agents in parallel
• span repos and environments from one place
• use local machines, worktrees, remote SSH, and cloud sessions in the same interface
• review diffs, commit changes, and manage PRs without leaving the workspace

That is a bigger product move than the marketing line suggests. Ryo Lu’s reaction picked up the same design goal, that the interface starts simple and exposes more tooling only when needed.

The release thread makes the scope explicit: agents can run locally, in a worktree, over remote SSH, and in the cloud. The blog post fills in the missing detail, which is handoff in both directions.

From cloud to local, a session can move down to the desktop when a developer wants to edit or test directly. From local to cloud, the same session can keep running after the user goes offline, which Cursor pitches for long tasks that would otherwise die when a laptop closes.

Cursor also says cloud agents now generate demos and screenshots for verification, reusing the same experience from cursor.com/agents inside the desktop app. Kevin Kern’s launch-day note focused on that seamless local and cloud switch, which is probably the part engineers will remember first.

The cleanest technical change lives in the official worktrees post. Cursor replaced older implicit worktree behavior with two explicit commands:

1. /worktree starts an isolated Git checkout for the rest of a chat.
2. /apply-worktree brings the result back to the main branch.
3. /best-of-n runs the same task across multiple models at once, each in its own worktree.
4. A parent agent comments on the different results, then lets you pick one or merge pieces from several implementations.
5. .cursor/worktrees.json can define project-specific setup, including dependency installs, env-file copying, and migrations.

That same thread includes a useful follow-up from Cursor engineer David Gomes: /worktree and /best-of-n now support multi-repo setups, which had been a practical blocker for teams splitting frontend and backend repos.

Shipper’s Every review lands roughly where the product evidence does. He argues the rebuild makes sense because agent orchestration is replacing plain AI pair programming, but he also says the current release still feels early and not yet strong enough to force a switch.

Cursor’s other notable choice is restraint. The launch post says the company is still carrying over the IDE features people actually use inside an agent-first shell, instead of pretending the editor no longer matters.

The list is concrete:

• file views for inspection
• go-to-definition in the editor with full LSP support
• a built-in browser that can open and navigate local websites
• marketplace plugins that extend agents with MCPs, skills, and subagents
• one-click installs, plus private team marketplaces

That is also why the rollout path is unusually lightweight for a product this different. Upgrade Cursor, open Cmd+Shift+P, launch Agents Window, and the old IDE is still there when you want it. Cursor is betting that agent orchestration gets adopted faster if it ships as an adjacent workspace instead of demanding that developers abandon the editor muscle memory they already have.