Cursor: PR Review in Agents, Parallel Build Plans, and Smart PR Splitting

PR Review, Parallel Build Plans, and Smart PR Splitting in Cursor 3.3

Cursor's May 7, 2026 release continues the Agents Window-first philosophy introduced in version 3.0, adding three meaningful workflow improvements for developers managing complex, multi-threaded work: an integrated PR review interface, parallel plan execution, and intelligent PR splitting.

PR Review Directly in the Agents Window

The most visible addition in this release is a fully native PR review interface. Previously, reviewing a pull request meant context-switching out of Cursor to GitHub or another tool. Now, the entire review workflow lives inside the editor.

The PR interface is organized across three tabs:

• Reviews tab — surfaces all inline review discussions and top-level comments in one view
• Commits tab — displays full commit history for the PR
• Changes tab — provides a file tree for navigating large diffs

Reviewer status, quick actions, and contextual discussions are all consolidated into this single panel. Cursor's goal with this integration is to let developers take a PR from creation through to merge without ever leaving the editor, reducing context-switching and keeping agents and humans in the same shared workspace.

Parallel Plan Execution

Cursor agents can now build plans in parallel. When an agent creates a multi-step implementation plan, the system analyzes the dependency graph of tasks and identifies components that can execute independently. Those independent steps are dispatched simultaneously via async subagents, while dependent steps continue to run in the correct sequential order.

This is a natural evolution of the /multitask capability introduced in earlier 3.x releases, now applied at the plan level rather than requiring manual orchestration from the user. The result is faster end-to-end execution for complex features that involve parallel workstreams — for example, writing tests while simultaneously refactoring a module, or scaffolding multiple independent files.

The /multitask command is also now available directly from the editor command palette, giving users a quick path to spawn async subagents without switching to the Agents Window.

Smart PR Splitting

Another new addition is a Split PRs quick action. When an agent has accumulated a large set of changes across multiple logical themes, this action analyzes the chat context to identify where the changes can be cleanly divided. The system defaults to independent PRs (unless dependencies between changesets are detected), creates a backup snapshot before splitting, and presents the proposed split plan for developer approval before executing.

This is particularly useful for long-running agent sessions where scope can drift — the splitting mechanism gives developers a structured way to clean up before merge, without manually cherry-picking or rebasing.

Pinned Skills as Quick Actions

Developers can now pin frequently used skills to the quick-action bar as pills — small, one-click launchers that surface at the top of the chat interface. This makes repetitive skill invocations (such as running a standard code review skill or triggering a specific automation) significantly faster.

Additional Improvements

Several smaller quality-of-life improvements shipped alongside the headlining features:

• Subagent configuration options are now accessible via the Explore panel
• General model name support added for subagent settings, giving more flexibility in model routing
• Enhanced undo/redo grouping for more predictable history navigation
• Visual stability improvements for long chat sessions
• MCP connection reliability improved with token cleanup on disconnect