Cursor released version 3.1 on April 13, 2026, with Canvas as the headline feature. The update changes what an Agent can return after finishing a task: not just code or Markdown, but a live React interface. Charts, tables, diff views, and custom logic can now be rendered directly on screen, without sending users to another window to build a visualization themselves.
Agent output moves beyond plain text
The main shift in 3.1 is the output format itself. Instead of ending with text-heavy results, an Agent can now package the result as an interactive React interface and keep it as a persistent artifact inside the Agents Window sidebar.
Canvas can stay visible alongside the terminal, browser, and version control panels. That changes the workflow in a simple way. Results from earlier Agent sessions do not disappear once the chat is closed, which makes them easier to revisit and reuse.
Charts and review views can combine more context in one place
Cursor described several concrete use cases. In earlier workflows, an Agent might collect data from Datadog, Databricks, and Sentry, then format it into Markdown tables for the user to interpret. With Canvas, the Agent can combine those sources and present them directly in a single chart.
The same idea applies to PR review. An Agent can group changes by importance and generate pseudocode for complex algorithms, instead of dumping a flat diff and leaving the sorting to the user. The result carries more information, but the reading load is lower.
Skill-based customization expands the Canvas system
Canvas also supports customization through Skills, and Cursor said the 3.1 architecture makes that ecosystem clearer. According to the release notes, Cursor Marketplace now includes the Docs Canvas Skill, which can generate interactive architecture diagrams for any code repository. Users can also write their own Skills to define what type of Canvas an Agent should produce.
Outside of Canvas, version 3.1 adds split view for running multiple Agents in parallel. Voice input has also changed: users can now hold Ctrl+M for batch STT transcription, giving them a different input workflow inside the editor.