Problems we're going to solve

The problem Cursor focuses on goes beyond simple code autocomplete, aiming to make code editing overall faster and more accurate. The key is to change not just model performance but also UX and infrastructure together.

The first thing needed is better context. Because there's too much information inside the editor—open files, similar code snippets, symbol connection relationships, lint output, execution traces, git history, typing history, external documentation, and more—Cursor is pursuing a method of gathering about 500K tokens per request and then compressing them with a custom reranker to leave only the most relevant 8k tokens.

The next challenge is "copilot for edits". While GitHub Copilot is useful for writing new code, making small but tedious modifications to existing code blocks—like renaming—still requires a lot of navigating, deleting, and typing. Solving this requires both a diff UX that's less disruptive while writing code, and model-side innovation that overcomes cost, latency, and intelligence issues together.

Cursor also aims for constrained, in-flow agents. This takes the form of a few-step agent working on a large codebase that searches, writes code, executes it, and periodically asks the user for confirmation. The initial goal is to make this work reliably within folders on the scale of hundreds of thousands of tokens, and if successful, expand to the entire codebase.

Beyond that, the following problems are also being addressed.

• Bug-finding: Continuously finding potential bugs in the background, or actively helping the user track down the cause while debugging.
• Larger edits: It should be possible to modify not just entire files but also directory-level changes, and changes should be shown in real time in a parseable form.
• Scale: Cursor has already indexed 1.4 billion vectors and 150,000 codebases, and this could grow 10x by the end of the year, requiring a faster indexing system.

Future ideas include previewing changes to be made within the next 15 minutes like Time warp, having the model deeply understand codebase concepts at the weights level through Understanding, providing assistance that explains code paths through Reader mode, reflecting abstracted modifications back into the original source through Pseudo-code mode, and even envisioning an experience where stack traces are automatically understood and fixed.

Finally, Cursor draws a line that this list is not a fixed roadmap but merely current priorities. Since the team uses the product themselves for more than 12 hours a day, ideas and priorities keep changing, but this is enough to show where Cursor is spending its brain cycles right now.