Cursor ditches VS Code, but not everyone is happy...

Two years ago, Cursor 1.0 was released as a VS Code fork for AI code autocompletion. Six months ago, Cursor 2.0 introduced an upgraded chat view capable of controlling the terminal to build entire features. Just recently, Cursor released version 3.0, which shifts focus from writing code to managing AI agents across multiple repositories, machines, and the cloud simultaneously, akin to an air traffic controller. Additionally, Cursor unveiled its new Composer 2 model, an in-house trained model that is claimed to be more intelligent than Claude Opus 4.6, based on internal benchmarks. A significant change in Cursor 3.0 is that it is no longer a VS Code fork, having been completely rewritten from scratch in Rust, which is beneficial for RAM usage. However, this new direction has sparked debate among users. The Composer 2 model, released about two weeks prior to this report, initially impressed with its benchmark numbers, suggesting intelligence exceeding Opus at lower cost and faster speed. However, transparency issues arose when it was discovered that Composer 2 is based on Moonshot's Kimmy K2 model, a detail not initially disclosed by Cursor. This was revealed when its model ID was found in Composer's metadata. Kimmy itself has faced accusations of training on Claude's outputs, occasionally identifying itself as "Claude." Cursor later apologized for the lack of transparency and released a technical report detailing their reinforcement learning approach with Kimmy. Despite the controversy, a model that is smart, fast, and cost-effective is considered valuable for the future of programming, particularly in a "zero code" environment. To facilitate the transition from code to agents, Cursor 3.0 features a new interface rewritten in Rust and TypeScript. While the traditional VS Code editor remains, the new interface is primarily designed for managing agents, allowing users to largely disregard their codebase due to the growing capabilities of coding agents. This environment integrates professional development tools, language servers, files, remote SSH, and the ability to run numerous agents anywhere. Critics have noted its similarity to OpenAI codecs. In a practical demonstration, the new Cursor was used to prototype a project called "Horse Tinder." The process began in "plan mode" to establish a basic architecture. During this, other tasks could be initiated in the background, such as developing a separate landing page, performing remote work via SSH on a cloud server, or switching to an entirely different project and launching an agent there. The system allows seamless parallel work with multiple agents from a single window. Agent status is monitored: a yellow dot indicates a need for human input, typically for permission to execute unsafe commands, while a blue dot signifies completed work ready for review. Within minutes, the system generated 13,000 lines of code. The interface allows checking Git history, opening a terminal, and inspecting code with a minimal file explorer. A notable feature is the built-in browser, which enables direct navigation to the app to interact with the final product. The demonstration showed a rendered SVG graphic, but the overall design was critiqued. Using "design mode," users can request changes by highlighting elements and instructing the AI to fix them, with the AI working in the background while users queue up further adjustments. This rapid iteration process was presented as a way to quickly develop complex applications.