Cursor v3 doubles code‑review speed, adds Composer and predictive editing
- Anysphere rolled out Cursor 3 on April 2, recasting its coding app as an agent workspace that runs parallel coding sessions across repos. - The release added a simpler diffs view, local-to-cloud agent handoff, worktrees, and Composer 2, Cursor’s in-house coding model built for speed. - Cursor is pushing beyond an AI editor toward “self-driving codebases,” where agents review, test, and ship code. (cursor.com)
Anysphere’s Cursor 3 turns Cursor from an AI-assisted editor into a workspace for managing multiple coding agents at once. (cursor.com) Cursor announced the release on April 2, 2026, and said the new interface was built from scratch around agents rather than files. The app can run agents locally, in git worktrees, in the cloud, and over remote Secure Shell sessions. (cursor.com 1) (cursor.com 2) The new Agents Window puts those sessions in one sidebar and lets developers move a task from cloud to local when they want to test edits on their own machine. Cursor says the reverse handoff also works, so a job can keep running in the cloud after a laptop closes. (cursor.com) Code review is part of that redesign. Cursor 3 added a new diffs view for reviewing changes, plus built-in staging, commits, and pull-request management inside the same window. (cursor.com 1) (cursor.com 2) The technical shift underneath the interface is codebase indexing, which scans source files so the agent can search a repository semantically instead of guessing from the open tab. Cursor’s docs say indexing runs automatically when a project opens and syncs periodically to pick up changes. (cursor.com 1) (cursor.com 2) Cursor is also leaning harder on its own models. Composer, introduced in October 2025, is Cursor’s in-house agent model; the company said it was trained for software engineering tasks and benchmarked at four times the generation speed of similar models. (cursor.com) In Cursor 3, Anysphere positioned Composer 2 as the fast model for quick iteration while agents move between local and cloud environments. The company also says users can switch among available models inside Cursor, tying model choice directly to the coding workflow. (cursor.com) (cursor.com) The release also added Design Mode, which lets a developer point at a live user-interface element in the browser and attach that element to an agent prompt. Cursor says that makes front-end feedback more precise by turning a screenshot target into part of the instruction. (cursor.com) A few weeks later, Cursor 3.2 extended that agent-first design with a `/multitask` command for async subagents and multi-root workspaces for cross-repo changes. That update suggests Cursor 3 was the start of a faster release cycle, not a one-off redesign. (cursor.com) Cursor’s own framing is broader than code completion. In its blog and changelog, the company describes a path toward “self-driving codebases,” with agents that review, test, and prepare changes to merge without leaving the development environment. (cursor.com) (cursor.com)
