Cursor 3.7 Adds Visual Prompting to Design Mode

On June 5, 2026, Anysphere released Cursor 3.7, introducing a spatial interaction model called Design Mode that allows developers to direct background agents using visual context instead of purely text-based chat. The update shifts the standard coding assistant workflow from a sentence-first approach to a visual-context-first model. By allowing users to point, draw, and narrate directly over running application interfaces, the IDE bridges the gap between design intention and code execution.

Spatial Interaction Mechanics

Design Mode introduces four primary methods for prompting agents visually against a rendered application interface.

• Point and Select: Clicking any UI element directly passes the element’s specific context, its underlying source code, and the surrounding layout structure to the agent.
• Drawing and Annotating: Developers can freeze a frame of the running application and draw bounding boxes, circles, or custom marks to define the exact spatial scope of a requested change.
• Multi-Select for Relationships: Users can highlight multiple distinct elements to describe relative changes, such as instructing the agent to match the padding of one component to another.
• Voice Integration: A persistent microphone overlay allows developers to narrate structural changes while the agent is still processing previous tasks. This enables a queued editing workflow without waiting for the hot-reload loop to finish executing.

Architecture and Optimization

The spatial features are optimized for Composer 2.5, an underlying model explicitly tuned for fast, targeted UI edits and rapid vibe coding loops. Design Mode is fully integrated with the newly released Cursor 3 environment and operates seamlessly within Canvases, the platform’s collaborative workspace for building interactive dashboards and reports.

The technical architecture relies on a tight hot-reloading loop. Agents apply code modifications in the background, and the application hot reloads natively inside the Cursor browser as soon as the file changes hit the disk. Because passing raw DOM data, spatial coordinates, and bounding box constraints to an agent requires massive context overhead, Cursor added a new interactive token usage report. This Canvas-level dashboard breaks down token consumption across system prompts, tool definitions, rules, and skills, giving developers a strict accounting of agent performance.

Enterprise Scale and Execution

Passing spatial data dramatically increases token throughput compared to standard text completions. Cursor supports this volume through the credit-based pricing model it adopted in 2025, avoiding strict per-request billing limits that restrict visual workflows.

The release aligns with a broader industry shift toward agent-first interfaces, where the IDE acts as a coordinator for multiple specialized subagents. Cursor currently reports active deployments across over 10,000 enterprises. At organizations operating at high utilization, such as NVIDIA with over 30,000 active developers on the platform, Cursor 3 and its background agents are credited with a three-fold increase in total committed code.

If you manage front-end deployment pipelines, Design Mode fundamentally alters the execution phase of the design-to-production loop. By treating the browser view as a visual diffing tool, you no longer need to manually translate graphical misalignments into text descriptions, allowing you to prompt the agent directly against the rendered output.