Cursor 2.0 is INSANE: First Look at Multi-Agents, New 4x faster Model, & Browser Mode

Key Concepts

• Cursor 2.0: The latest version of the AI-powered code editor, introducing significant new features and improvements.
• Agent Mode: A new mode in Cursor where AI agents perform most of the coding tasks, with the user primarily reviewing the results.
• Work Trees: A feature that allows Cursor to attempt to solve a task multiple times, presenting different versions for the user to choose from, simplifying Git management.
• Browser Tab Feature: An integrated browser within Cursor that allows users to interact with web pages directly in the editor, including element selection and code referencing.
• Composer 1: A new, highly performant coding model introduced in Cursor 2.0, known for its speed and cost-effectiveness.
• Use Multiple Modes: A feature enabling the use of several AI models simultaneously to tackle a task, generating multiple solutions for comparison.
• Parallel Task Execution: The ability to run multiple AI agents or models concurrently to speed up development and task completion.
• Browser Agentic Browsers: The browser feature in Cursor that can be used for automating workflows and interacting with web pages.
• Custom Commands: The ability to create and share custom commands with team members within Cursor.
• Performance Improvements: Enhancements made to Cursor to improve its speed and reduce hanging issues.
• Plan Mode in Background: The capability to initiate planning and task execution for multiple projects simultaneously in the background.
• Claude Code: An alternative AI coding tool that is generally more cost-effective but requires more manual setup for advanced features.
• Shipkit.ai: A platform offering AI application templates, courses, and resources for developers.

New Features and High-Level Changes in Cursor 2.0

Cursor 2.0 introduces a significant overhaul with several key features designed to enhance developer productivity and AI integration. The most noticeable change is the introduction of Agent Mode, which shifts the primary coding responsibility to AI agents. Users can initiate tasks, and the agents will work in the background, with the user's role becoming one of review and validation. This mode supports both local and cloud-based execution of these background workers.

A core component of Agent Mode is the Work Tree functionality. This feature allows Cursor to generate multiple solutions or versions of code for a given task. The AI attempts to solve the problem iteratively, and the user can then select the preferred version to apply to their codebase. This process is designed to abstract away complex Git operations, simplifying the management of code changes.

Enhanced Interaction and Browser Integration

Voice Input and Whisper Flow

Cursor 2.0 incorporates a new mic button for voice input. Users can speak commands to interact with their code, such as specifying file modifications. The transcript notes a comparison with Whisper Flow, another tool, suggesting that while Cursor's voice input is functional, Whisper Flow currently demonstrates better codebase understanding and automatic file referencing. A limitation mentioned is that voice input can only be initiated at the start of a new chat session.

Integrated Browser Tab

A significant addition is the browser tab feature, which opens a new browser instance within Cursor. This allows for seamless interaction with web pages. Users can:

• Open a new browser instance to navigate to any URL.
• Bring an existing browser tab into the Cursor editor.
• Select elements on a web page (e.g., an H1 tag) and have Cursor automatically reference the corresponding code in the editor. This enables direct modification of website elements by speaking commands like "change this to say super fast and affordable." This feature is highlighted as particularly useful for debugging UI issues and making quick visual changes.

The browser feature also allows for the creation of browser agentic workflows. Users can perform a series of actions on a webpage once, and then instruct Cursor to record these actions and save them as a set of instructions for future automation. This is beneficial for repetitive tasks like filling out timesheets.

However, a key limitation of the browser tool is its inability to resize the editor window, which hinders its effectiveness for tasks like making a website mobile-responsive by analyzing screenshots at different breakpoints. Additionally, websites requiring user login will necessitate manual authentication within the new browser instance.

New Models and Parallel Processing

Composer 1 Model

Cursor 2.0 introduces Composer 1, a new coding model that is described as an "absolute speed demon." Initial tests suggest it performs comparably to Sonnet models in terms of code quality and is significantly faster. The blog post from Cursor indicates Composer 1 is nearly as intelligent as top-tier frontier models like Claude 4.5 Sonnet.

"Use Multiple Modes" and Parallel Task Execution

The "Use Multiple Modes" feature allows users to select and run several AI models concurrently on a single task. This enables parallel task execution, where different models attempt to solve the same problem, providing multiple variations for the user to choose from. While this consumes more tokens, it offers a wider range of solutions.

A practical demonstration involved a task to grant admin users unlimited usage in a startup project. Two models, Composer 1 and Sonnet 4.5, were run in parallel. Composer 1 completed the task significantly faster than Sonnet 4.5, showcasing its speed advantage. The results from both models were functionally identical, but Composer 1 was four times faster and three times cheaper.

Cost and Performance Comparisons

Composer 1 vs. Claude 4.5 Sonnet

A detailed cost analysis revealed that Composer 1 is substantially more cost-effective than Claude 4.5 Sonnet. For a small to medium task, Composer 1 cost approximately $1 per 1 million tokens, while Claude 4.5 Sonnet cost around 33 cents per 1 million tokens. This translates to Composer 1 being three times cheaper and four times faster for comparable results.

Composer Model Development Insights

The Cursor team's blog post on Composer 1 highlights its development process. It started with a base open model and was refined through reinforcement learning using real-world coding development environments. Key improvements include:

• Enhanced Intelligence: Achieving intelligence levels comparable to frontier models.
• Parallel Tool Usage: Optimizing the model to use tools in parallel, rather than sequentially, for increased speed.
• Increased Search and File Read Operations: Composer 1 utilizes more search and file read tool calls, leading to more informed and intelligent solutions by analyzing the codebase more thoroughly.

Other Notable Features and Improvements

• UI Enhancements: The user interface has been refined for better usability, including more togglable elements for code reviews.
• Custom Commands Sharing: The ability to create and share custom commands with team members is now available, facilitating consistent development practices.
• Performance Optimizations: Efforts have been made to improve Cursor's performance and reduce instances of hanging.
• Background Plan Mode: Users can now initiate planning and task execution for multiple projects simultaneously in the background.
• Notifications: New notification features allow users to be alerted when background agents complete tasks, with options for menu bar indicators and sound alerts.

Recommendation: Cursor 2.0 vs. Claude Code

The presenter offers a nuanced recommendation between Cursor 2.0 and Claude Code:

• Claude Code: Generally more cost-effective, especially for users who are comfortable with manual setup of advanced features like work trees and parallel task execution. It is recommended for more technical engineers who are cost-sensitive and prefer working with the terminal.
• Cursor 2.0: Offers a more user-friendly and integrated experience with features like Agent Mode, Work Trees, and the Browser Tab readily available. It is recommended for "vibe coders" who prioritize ease of use and are less concerned about token costs. The presenter notes that while Claude Code has had these features for a while, Cursor 2.0 makes them significantly more accessible and beautiful in their implementation.

The presenter plans to continue using both tools, leveraging Claude Code for its cost-effectiveness in generating code projects and Composer 1 within Cursor for its speed and efficiency. The final decision on which tool is "best" for cost-effectiveness is still pending further testing.

For most users, Cursor 2.0 is recommended due to its superior UI and integrated features, despite a slightly higher cost. Power users who are highly cost-sensitive might find Claude Code a better fit.

Conclusion

Cursor 2.0 represents a significant leap forward in AI-powered code development. The introduction of Agent Mode, Work Trees, the integrated browser, and the highly performant Composer 1 model offer substantial improvements in speed, efficiency, and user experience. While Claude Code remains a strong contender for cost-conscious developers, Cursor 2.0's seamless integration and advanced features make it a compelling choice for a broad range of developers.