The world of AI-powered coding just got a serious jolt. Cursor, the AI-first editor that started as a popular fork of VS Code, announced version 2.0. This isn’t just an incremental update with bug fixes; it’s a fundamental rethinking of what a code editor should be, centered on a new multi-agent interface and the debut of the company’s first-ever in-house coding model.

If you’ve used AI assistants like GitHub Copilot, you’re used to the “pair programmer” model—an AI that sits beside you, offering suggestions and completing lines. Cursor 2.0 is making a bold bet on a different future: one where you’re not just a programmer with an assistant, but the manager of a whole team of AI agents.

“The first thing you’ll notice when you open Cursor is our new interface,” the Cursor team writes in their announcement post. “It’s more focused and designed from the ground up to be centered around agents rather than files.“

This is the core philosophical shift. For decades, developers have organized their work around a tree of files and folders. Cursor 2.0 aims to abstract that away. The new UI is built to let you “focus on the outcomes you want while agents take care of the details.“

In practice, this means your primary interaction is no longer just typing in a text file. It’s describing a feature, a bug fix, or a refactor in plain English. You then dispatch AI agents to execute that plan. For those who get nervous about the AI “black box,” Cursor says you can “still easily open files in the new layout or switch back to the classic IDE.”

Meet the (parallel) team

The headline feature of this new interface is its multi-agent capability. This is where things get really interesting.

Cursor 2.0 allows you to run multiple AI agents in parallel, and crucially, without them interfering with each other. Any developer who has tried to run two large-scale AI tasks at once knows the chaos that can ensue as they overwrite each other’s work or tangle up dependencies.

Cursor solves this by using a clever sandboxing technique powered by Git worktrees or remote machines. Each agent gets its own clean, isolated copy of the codebase to work in.

This opens up two powerful new workflows:

1. Divide and conquer: You can assign different tasks to different agents. Imagine dispatching one agent to refactor your API endpoints while a second agent simultaneously writes the corresponding unit tests.
2. The “model bake-off”: You can assign the exact same problem to multiple different models at the same time. For instance, you could ask Cursor’s new Composer, GPT-5, and Claude 4.5 Sonnet to all “fix this bug.” Up to eight agents can work in parallel. You can then review all three proposed solutions and pick the best one. As Cursor puts it, this “significantly improves the quality of the final output,” especially for complex tasks.

‘Composer’: an in-house model built for speed

Running all these agents is great, but it can be slow. Waiting 60 seconds for an AI to “think” is a notorious “flow” killer for developers. To solve this, Cursor 2.0 is launching Composer, its first-ever proprietary model.

Composer is described as a “frontier model for low-latency agentic coding.” The company claims it’s four times faster than “similarly intelligent models,” with most conversational turns completing in under 30 seconds.

This isn’t just a generic model. It was trained specifically for the kind of work agents do, using tools like codebase-wide semantic search. This gives it a deep understanding of large, complex codebases—a common weak point for general-purpose models that might not understand how all your project’s files connect.

Based on early benchmarks, Composer appears to make a strategic trade-off: it may not beat the absolute “smartest” (and slowest) frontier models in raw intelligence, but it’s dramatically faster, aiming for the sweet spot of being “fast enough to feel interactive.”

Solving the problems AI created

In a moment of insightful self-awareness, the Cursor team notes that as AI gets better at writing code, it creates two new bottlenecks for the human developer: reviewing and testing all that new code.

Cursor 2.0 introduces features to tackle both.

First, a new code review interface makes it much easier to see the changes an agent has made across multiple files at once, rather than forcing you to jump between dozens of tabs.

Second, and perhaps more impressively, Cursor 2.0 includes a native browser tool that is now generally available. This tool allows an AI agent to test its own work. The agent can literally open a browser within your editor, interact with your web app’s UI, click buttons, fill out forms, and verify that its changes actually work. If it finds a bug, it can iterate on its own solution until the tests pass, closing the loop on the development process.

This update pushes Cursor far beyond its origins as a “VS Code fork with AI.” It’s a comprehensive platform built on a new “intent-driven” paradigm, betting that the future of software development lies in managing a team of AI agents, not just pairing with one.