Cursor 3.7 Makes Agent Context Easier to Inspect

What this is about

Cursor released version 3.7 on June 4, 2026. The most notable change is not a larger model or another autopilot feature, but more visibility: Cursor can now show an agent's context usage as an interactive report inside a canvas.

That may sound dry, but it matters for real software work. Many coding agents fail not because the model is generally too weak, but because too much, too little, or the wrong context reaches the prompt. Cursor lists system prompts, tool definitions, rules, skills, and other sources among the context components. In many tools, those parts remain mostly invisible.

What Cursor 3.7 actually does

Cursor 3.7 brings two connected changes. First, Design Mode is now available in canvases. Users can select and annotate UI elements directly in a canvas instead of describing a visual change only in text.

Second, the new Context Usage Report shows how the agent input is assembled. According to Cursor, the report breaks down where tokens go across the system prompt, tool definitions, rules, skills, and more. Because the report itself appears as a canvas, users can ask follow-up questions or use the embedded Debug with Agent button to look for ways to reduce context in a new conversation.

Why it matters

Context is budget, risk, and quality control for coding agents. Cursor's own documentation describes context as a mix of intent context and state context: what the user wants and what the current project state is. If relevant state is missing, hallucinations and unsuitable edits become more likely. If too much irrelevant state is carried along, costs rise and the actual task can disappear in the noise.

For teams, this transparency becomes especially valuable when rules, AGENTS.md files, MCP tools, and many project files meet in the same session. Cursor documents that rules can be included as reusable context at the start of the prompt. That is useful, but over weeks it can also become hidden prompt weight.

The release is therefore less a convenience feature than a small step toward inspectable coding agents. Teams using agents in production repositories need answers to simple questions: Why did the agent read this file? Which rule was active? Which tool definitions expanded the context? Cursor 3.7 tries to move those questions closer to daily work.

In plain language

Imagine packing a suitcase for a three-day trip. If you cannot see what is already inside, you may pack three jackets and forget the toothbrush. The Context Usage Report is like opening the suitcase: you can see what takes up space, what is missing, and what can come out.

Design Mode is the second part of the same idea. Instead of saying, "change the thing in the upper right that kind of looks like a button," you can point to it directly. That reduces misunderstanding, especially in visual work.

A practical example

A team asks Cursor to rebuild an internal admin interface. The session has 180,000 tokens of context: 45,000 tokens come from project rules, 30,000 from tool descriptions, 70,000 from read files, and the rest from chat history and error messages.

In the Context Usage Report, the team sees that three old rules still accompany every request even though they only applied to a past experiment. Removing them cuts context by 25,000 tokens per task. Across 200 agent runs per week, that is not only cheaper; it is also easier to review because fewer irrelevant instructions travel with the agent.

Scope and limits

• The report shows how context is assembled, but it does not prove exactly which information the model used for a decision.
• Visibility does not replace code review. An agent can still write incorrect changes with clean context.
• The feature helps most for teams that use Cursor heavily. For occasional one-off tasks, the value is smaller.

The important point is that Cursor 3.7 does not solve the broader governance problem around coding agents. It makes one part of the black box more visible. That is useful, but it is not a guarantee of security, privacy, or reproducible results.

SEO & GEO keywords

Cursor 3.7, Cursor Canvas, Context Usage Report, AI coding agent, coding agents, token budget, AGENTS.md, Cursor Rules, developer tools, AI code review, MCP, software engineering