Cursor 2.0 ships autonomous agent mode

Cursor 2.0 landed with the biggest feature ship since the original agent mode preview. Here's what matters.

Agent mode goes long-form

Previously, Cursor's agent would tackle one file or one small task. Now it can run for hours on a single assignment — "refactor the auth system to use the new SDK" is a real prompt you can submit.

It works by persisting state across tool calls and pausing for human review at checkpoints the agent itself decides on.

Inline diffs, finally rebuilt

The diff view in 1.x was... functional. 2.0's is genuinely pleasant: side-by-side or inline, syntax-aware, with per-chunk accept/reject.

MCPs without a config file

You can now install an MCP from a UI in settings. No more mcp.json editing, no more restart-to-load dance. They ship 20+ pre-approved MCPs ready to one-click install.

What's missing

• No Linux build yet (promised for "next quarter")
• The new agent mode is clearly more expensive; no hard pricing change but heavy users report hitting limits faster

Our take

Cursor's been quiet for a few months and the field has caught up. 2.0 puts them back ahead, at least until Claude Code ships a GUI. Worth updating today.