Cursor released version 3.1 on April 15-16, 2026 with a new feature called Canvas that changes what AI agents produce as output. Instead of generating only code or markdown text, agents running in Cursor’s Agents Window can now create interactive React-based interfaces containing charts, tables, diff views, diagrams, and custom logic, according to Cursor’s official blog.
Canvas outputs are persistent artifacts that live alongside the terminal, browser, and source control in the Agents Window. They are rendered using a React-based UI library with first-party components, and developers can extend them through custom Skills available on the Cursor Marketplace.
Incident Response With Multi-Source Data Visualization
Cursor described several internal use cases that drove the feature. For incident response, the team connects Datadog, Databricks, and Sentry through MCP integrations. Before Canvas, agents compiled time-series data from these sources into markdown tables that were difficult to interpret and required manual visualization steps. Now agents consolidate multi-source observability data directly into a single interactive chart within a Canvas.
For PR reviews, Canvas lets agents logically group changes by importance and present a structured interface for exploring the change set. Agents can write pseudocode representations for complex algorithms rather than laying out all diffs with equal weight. As BlockBeats reported via KuCoin News, the core improvement is information density: data that previously required manual post-processing now arrives in a reviewable visual format.
Cursor’s Own Model Rollouts as a Test Case
The Cursor team used Canvas to build their eval analysis workflow. Engineers previously inspected request IDs one at a time to identify failure patterns during model releases. They considered building a separate web application to automate the process. Instead, they created a Skill within Cursor that allows agents to read rollout data, group failures by cause, and generate an interactive analytical interface. The team reported that this workflow helped them release two new models with significantly less effort.
The team also adapted Andrej Karpathy’s autoresearch framework to run optimization experiments, using Canvas to visualize research progress and active hypotheses while experiments run.
Agent Output Becomes Interactive Artifact
Canvas is part of a broader pattern at Cursor alongside Design Mode and upgraded voice input, all aimed at increasing the information bandwidth between humans and AI agents. The shift is from “AI produces code, human reads code” toward “AI produces interactive artifacts, human reviews outcomes.” A developer reviewing an agent’s work can interact with a live chart of analyzed data rather than reading through the agent’s code to understand what it found.
Canvas is available now in Cursor 3.1. The Docs Canvas Skill on the Cursor Marketplace lets agents generate interactive architecture diagrams for any code repository.