Cursor

You can now assign a different model to each Agent in separate tabs and editors. Forking an Agent preserves the selected model.

Usage stats are now visible in Chat. You can see your usage either all time, or when you’ve exceeded 50% of your plan’s quota. This summary can be toggled in Settings. Team and Enterprise users will have access to the usage summary in September. 

Compact mode hides tool icons, collapses diffs by default, and auto-hides the input when idle. This can be useful for long sessions with many tool calls, terminals, and diffs. You can toggle compact mode on from settings.

Agents can now be used directly inside GitHub pull requests. Tag @Cursor and the Agent will read the prompt, apply the fix, and push a commit. Background Agent will keep you updated on status with a comment and todos. https://cdn.sanity.io/images/2hv88549/production/2785959df1ef39ad4795ec260b2f89fc22f2ece7-1854x978.png?auto=format

You can now inspect and manage all Agents from the left sidebar, both foreground and background. Peek into remote machines to see Background Agent working by clicking on one in the sidebar.

Sending messages while Cursor is working now does a better job of steering the agent. Messages will run at the next ideal time, usually after a tool call, versus waiting until the generation completes. ⌥+Enter (Alt+Enter on Windows) will queue messages like before and ⌘+Enter (Ctrl+Enter on Windows) will interrupt the agent to send the message immediately. The default behavior can be changed in Cursor Settings → Chat → Queue messages.

Agents are significantly improved when operating across large codebases. Context selection is more relevant, token use is more efficient, and edit quality is higher: Read file: Now reads full files when appropriate and the 2MB cap is removed.List: Can now explore full directory trees in one call, with metadata like file counts and types.Grep: Improved matching with less noise.Codebase Search: Better ranking and indexing, leading to more relevant context.Web Search: Tuned for smaller, more relevant responses using a lightweight native model.

We’ve put a lot of effort into improving Background Agents and they are now twice as fast to start.