Cursor 2.0 and Composer

What happened

On 29 October 2025, Cursor released Cursor 2.0, and with it Composer, the company's first coding model ^[1]. The release is not only a model update: it also changes how the editor is organized and adds new ways to run and check the work of coding agents ^[1].

The model

Cursor describes Composer as a frontier model that is four times faster than similarly intelligent models ^[1]. It is built for low-latency agentic coding, and Cursor says most turns complete in under 30 seconds ^[1]. That speed matters because an agentic coding session is a loop: the model reads code, proposes a change, runs a tool, reads the result, and continues. A shorter loop makes it practical to let the agent work through several steps instead of waiting for a single long response.

Composer was trained with tools, including codebase-wide semantic search ^[1]. Semantic search lets a model find code by meaning rather than only by exact names, so it can gather relevant files across a large repository. Cursor says this makes the model much better in large codebases ^[1]. The point is not that the model memorizes a repository; it is that the model learns to use search as part of its work.

Interface and parallelism

The new interface is designed around agents rather than files, while the classic IDE layout remains available ^[1]. This is a shift in the unit of work. In a file-centered editor, the human opens a file and edits it. In an agent-centered interface, the human describes a task and watches or directs one or more agents that act on the codebase.

Cursor 2.0 can run many agents in parallel without interference, powered by git worktrees or remote machines ^[1]. Git worktrees let separate working directories share one repository, so parallel agents do not have to edit the same files at the same time; remote machines provide another way to isolate runs. Cursor also found that having multiple models attempt the same problem and picking the best result significantly improves output on harder tasks ^[1]. This is a form of best-of-N sampling: instead of trusting one attempt, the system generates several candidates and selects among them. It costs more compute, but it can raise quality on tasks where a single pass is unreliable.

Review and testing

The release names two emerging bottlenecks: reviewing code and testing changes ^[1]. Generating code is only part of the job; someone still has to read the diff, and the change still has to be run. To help with the second bottleneck, Cursor adds a native browser tool so that Cursor can test its work and iterate ^[1]. A browser tool lets an agent open a web application, interact with it, observe the result, and adjust the code. That closes a loop that otherwise requires a human to switch to a browser and report back.

Why it matters

Cursor 2.0 shows where AI coding tools are moving: from suggesting lines inside a file to coordinating agents, isolating their work, comparing their outputs, and giving them tools to check the result ^[1]. The release does not remove the human from the process. Reviewing code and testing changes remain the bottlenecks Cursor itself names ^[1]. But by making agents faster, parallel, and able to test in a browser, Cursor 2.0 tries to make the agentic workflow practical rather than experimental.